2026腾讯游戏安全竞赛Android初赛

2026腾讯游戏安全竞赛Android初赛 wp

初赛对godot的理解不够，以前学习的主要是UE4相关，godot还是第一次搞，godot也有类似ue4那种dump，决赛才发现.

题目附件：下载压缩包

主要目标

• 正算法：根据屏幕左上角的随机token生成右上角的flag
• 逆算法：根据屏幕右上角的flag反推左上角的token
• 按实现的算法数量和完成度计分

到达绿色方块也可获得flag

Flag获取 [Frida Hook碰撞]

更简单的方法是把属性dump出来直接去触发碰撞，我这样调得出来的效率太低了

补充patch掉校验 主动触发 然后重新编译打包 失败会闪退 估计是得用原来的引擎或版本不对?

这个其实应该放在后面的，我是解完包，逆完算法在开始的，还好搞了，不对不知道，自己逆向的算法搞错了

要想获得flag 无非几种常见思路

1. 主动调用process
2. 主动撞箱子，飞上去
3. 主动触发装箱函数

由于初次解出godot对感觉没有ue4那么有规律，也不知道怎么看字符串，对照方法，只能硬调

当前方案是用 Frida 强制 Trigger2 碰撞,让真 flag 显示在 Label2 UI 上，decrypted/decompiled/trigger.gd 是挂在 Trigger1 和 Trigger2 两个 Area3D 节点上的共用脚本。简化后的_process逻辑:

func _process(delta):
    # ...
    var body = get_overlapping_bodies()
    if body.size() > 0:
        if str(get_path()) == "/root/TownScene/Trigger1":
            label.text = "flag{sec2026_PART0_example}"     
        elif str(get_path()) == "/root/TownScene/Trigger2":
            if not flag2Triggered:
                flag2Triggered = true
                # 读 Label.text 里的Token: xxxxxxxx, 调 GameExtension.Process 算真 flag
                label.text = "flag{" + FLAG_PREFIX + flag1 + "}  "

Trigger2 是屋顶的绿色方块,是一个可穿过的 Area3D 触发器(应该不是实体),只是 Y 高度超出车能到达的范围,所以 get_overlapping_bodies() 永远返回空数组,Process永远不被调用,flag 永远不显示。

Part1 的 20 分要求让 Trigger2 触发一次,使真 flag 出现在右上角 Label2 UI 上

不改 APK (之前试过改 trigger.gdc 再打包,改出来游戏直接崩,不是签名问题就是字节码格式问题),改用 Frida 在运行期 hook Area3D::get_overlapping_bodies,在 onLeave 里调用 Array::resize(out_array, 1) 强制把返回的空数组 resize 成size=1,trigger.gd 的 body.size() > 0 判断就过了,Trigger2 的分支自然执行,flag 直接写到 Label2,不需要替换字符串也不需要走 UI hack。

get_overlapping_bodies：获取当前这个 Area3D 正在重叠，碰撞到的物体列表

两个难点:

1. 区分 Trigger1 vs Trigger2。Trigger1 每帧都能触发 (玩家若撞到它),它会写flag1可能覆盖 Trigger2 的flag。必须只给 Trigger2 force,不动 Trigger1。
2. 找到真正被 GDScript 调用的 Area3D::get_overlapping_bodies 函数地址。libgodot_android.so 没有符号,得自己定位。

由于不太懂godot,一开始找错了函数 (Area2D 不是 Area3D)

第一次用 find_overlapping_bodies.py扫 libgodot_android.so,找 “get_overlapping_bodies”
字符串,再从 _bind_methods 的 ADRP+ADD pair 找到目标函数指针,结果定位到:

sub_27AF370  (vaddr 0x27af370)

反编译看起来非常像 get_overlapping_bodies:sub_3957654 构造 Array 检查monitoring flag 遍历链表 sub_395126Cresize 返回。但里面Array::set_typed 传的类名字符串是Node2D:

Node2D = "Node2D";               // Area2D的
sub_3C5FA18(&Node2D, &obj_);     // 构造 StringName
sub_3956F20(a2, 24, &obj_, ...); // set_typed(OBJECT, "Node2D")

把 0x27af370 塞进 Frida hook 后,调用次数一直是 0,但 Array::resize 探针每 2 秒能记到 20+ 次调用,证明 Frida 能 hook libgodot_android.so,就是这个函数从来没被执行过 , 它是 Area2D::get_overlapping_bodies,游戏里没 Area2D 节点,所以根本不跑。find_overlapping_bodies.py 脚本的启发式把 Area2D 和 Area3D 弄反了。

脚本就不贴了，没啥用，思想就是

就是扫 libgodot_android.so 找 godot::Area3D::get_overlapping_bodies的 vaddr 偏移。

在 Godot 的 Area3D::_bind_methods() 里, 会有一段ClassDB::bind_method(D_METHOD(“get_overlapping_bodies”), &Area3D::get_overlapping_bodies);编译后的 ARM64 汇编是:

ADRP Xa, <"get_overlapping_bodies" string page>   加载字符串地址
ADD  Xa, Xa, :lo12:<string>
...
ADRP Xb, <&Area3D::get_overlapping_bodies page>   加载函数指针
ADD  Xb, Xb, :lo12:<function>
BL  <ClassDB::bind_method>

这两对 ADRP+ADD 之间通常只差几条指令.我们扫 libgodot_android.so 的 .rodata 找 “get_overlapping_bodies\0” 字符串,

再扫 .text 里引用这个地址的 ADRP+ADD, 然后在引用点附近找第二对 ADRP+ADD,

目标就是 Area3D::get_overlapping_bodies.

注: Area2D 也有同名方法, 所以会找到 2 个候选, 选 arm64 的第二个大概率是 Area3D.

最后是用 LR反向追踪找到真函数

既然 Array::resize (sub_395126C，这里是被去掉符号了) 的 hook 能生效,就在它的 onEnter里读 this.context.lr(= caller’s next PC = BL resize 指令的下 4 字节)反向定位所有调 resize 的 call site,采样 10 秒:

为什么要 hook 它，Area3D::get_overlapping_bodies 内部会两次调用 Array::resize:

• 进循环前 resize(a2, 初始估计 count) 预分配
• 遍历链表收集完 bodies 后 resize(a2, 实际 count) 收尾

所以 trigger.gd 每帧每个 trigger 跑一次 get_overlapping_bodies,就会触发 2 次 Array::resize 调用。hook 这个函数等于间接监视到了 get_overlapping_bodies的内部活动,尽管我们还不知道 get_overlapping_bodies 的地址。

什么意思嘞，可以看效果解释

Interceptor.attach(resizeAddr, {
    onEnter: function(args) {
        const lr = this.context.lr;
        const callSiteOff = lr.sub(4).sub(mod.base).toInt32();
        callSites.set(callSiteOff, (callSites.get(callSiteOff) || 0) + 1);
    }
});

开车 10 秒后,top hits:

this.context.lr LR 寄存器BL 指令的下 4 字节，LR 是 Link Register,ARM64 里叫 X30,专门存函数返回地址，ARM64 的函数调用指令是 BL target,执行时 CPU 做两件事:

LR = PC + 4    ; 把"下一条指令地址"存到 LR (所以调用方返回时用)
PC = target    ; 跳转到被调函数

举个例子,假设 get_overlapping_bodies 里有一条指令:

0x25fa790:  BL  sub_395126C    ← 调 Array::resize
0x25fa794:  MOV X0, X1         ← BL 的"下一条"

当 CPU 执行到 0x25fa790 这条 BL 时:

• LR = 0x25fa794 (= BL 指令的地址 + 4 字节,因为 ARM64 每条指令固定 4 字节)
• 然后跳到 sub_395126C 开始执行

此时 Frida 的 Interceptor.attach(sub_395126C) 的 onEnter 会被触发。在 onEnter 里 this.context.lr 就是当时的 LR 值 = 0x25fa794。 想知道谁调了我? 做一个减法:

call_site = lr - 4 = 0x25fa794 - 4 = 0x25fa790

0x25fa790 就是那条 BL sub_395126C 的地址,也就是call site(调用点)。它所在的函数就是调用者。

1392≈ 60 FPS × 2 triggers × 10 秒 ，完全对得上两个 Trigger 每帧各调一次。两个 call site 都在同一函数 sub_25FA6F4 0x25fa6f4 内 (一次用于初始 resize,一次用于最终 resize)。反编译这个函数:

int64 Area3D::get_overlapping_bodies(Area3D *a1@X0, Array *a2@X8) {
    sub_3957654(a2);                              // Array::Array()
    // ...
    Node3D = "Node3D";                            // 真正的 Node3D
    sub_3C5FA18(&Node3D, &obj__1);
    sub_3956F20(a2, 24, &obj__1, &Node3D);        // set_typed(OBJECT, "Node3D")
    if ((*(_BYTE *)(a1 + 1400) & 1) == 0)         // monitoring flag @ +1400
        return error();
    sub_395126C(a2, *(_DWORD *)(a1 + 1444));      // 初始 resize,call site #1
    v5 = *(_QWORD **)(a1 + 1424);                 // 链表头
    // ... 遍历链表收集 bodies ...
    return sub_395126C(a2, v6);                   // 最终 resize,call site #2
}

识别 Trigger2 的 this 指针，知道了函数地址,还得区分 Trigger1 和 Trigger2。

get_overlapping_bodies 这个函数是所有 Area3D 都用的。Trigger1 和 Trigger2 都是 Area3D如果不区分很可能会看不出效果

解决办法是看 trigger.gd 的代码,obj.Process() 这行只在 Trigger2 的分支里出现。Trigger1 分支压根不调 Process

办法借助 GameExtension::Process 的调用时机。trigger.gd 里 obj.Process() 只在 Trigger2 分支里被调:

elif str(get_path()) == "/root/TownScene/Trigger2":
    if not flag2Triggered:
        # ...
        var flag1 = obj.Process(xor_enc(str(label1.text).substr(7)))

所以:

1. 每次 Area3D::get_overlapping_bodies 进入时记下 state.lastArea3D = args[0]
2. 每次 GameExtension::Process 进入时,当前调用它的就是 Trigger2,而此时 lastArea3D就是 Trigger2 的 this 指针 (因为 Trigger2 的 _process先调 get_overlapping_bodies 再调 Process,中间没别的 Area3D 调用)
3. 缓存 state.trigger2Ptr = state.lastArea3D,从此只给 Trigger2 force,Trigger1 和其他 Area3D完全不动

(trigger2Ptr == null)对所有 Area3D force,确保 Trigger2 那帧能进Process 分支,把自己的 this 指针”暴露”出来。一旦暴露,立刻切换到只 force Trigger2模式 ，Trigger1 及其他 Area3D 的 get_overlapping_bodies 不再被干扰。

流程时序 (两种可能的场景树顺序都 OK):

• Trigger1._process 先跑: 被 force, 写假 flag 到 Label2 (1 帧,临时)
• Trigger2._process 后跑: 被 force, 写真 flag 到 Label2, 调 Process,Process 的 hook 捕获 lastArea3D = Trigger2.this
• 下一帧: Trigger1._process 不再被 force, body.size() == 0, 不写 Label2,真 flag 保留在 Label2 上

反过来 (Trigger2 先跑、Trigger1 后跑) 也可以,因为一旦 Process 被调、trigger2Ptr被设,后续 Trigger1 不再被 force,Trigger1 分支不执行,真 flag 不被覆盖。

const GET_OVERLAPPING_BODIES_OFFSET = 0x25fa6f4;
const ARRAY_RESIZE_OFFSET           = 0x395126c;

const PROCESS_OFFSET = 0x4e548;

const FLAG_PREFIX = "sec2026_PART1_";

rpc.exports = {
    getState: function() { return JSON.stringify(state); },
};

// Godot Array 内部布局读取

function getArraySize(arrayPtr) {
    try {
        const arrayPrivate = arrayPtr.readPointer();
        if (arrayPrivate.isNull()) return -1;
        const dataPtr = arrayPrivate.add(16).readPointer();
        if (dataPtr.isNull()) return 0;
        return dataPtr.sub(8).readU64().toNumber();
    } catch (e) {
        return -1;
    }
}

// 状态
const state = {
    lastArea3D: null,        // 最近一次 get_overlapping_bodies 的 this 指针 (string)
    trigger2Ptr: null,       // Process 第一次被调时确定的 Trigger2 的 this 指针
    realFlag: null,          // 从 Process 返回值读到的真 flag
    forceCount: 0,           // 统计 force 了多少次 (debug)
};

// Hook 安装
let arrayResize = null;

function installHooks() {
    const godot = Process.findModuleByName("libgodot_android.so");
    const sec   = Process.findModuleByName("libsec2026.so");
    if (!godot || !sec) {
        setTimeout(installHooks, 300);
        return;
    }
    console.log(`[+] libgodot_android.so base = ${godot.base}`);
    console.log(`[+] libsec2026.so       base = ${sec.base}`);

    const gobAddr    = godot.base.add(GET_OVERLAPPING_BODIES_OFFSET);
    const resizeAddr = godot.base.add(ARRAY_RESIZE_OFFSET);
    const procAddr   = sec.base.add(PROCESS_OFFSET);
    console.log(`[+] Area3D::get_overlapping_bodies @ ${gobAddr}`);
    console.log(`[+] Array::resize                  @ ${resizeAddr}`);
    console.log(`[+] GameExtension::Process         @ ${procAddr}`);

    // dump first 16 bytes of the hooked region to verify we're in code
    try {
        const b = gobAddr.readByteArray(16);
        const bv = new Uint8Array(b);
        let hex = "";
        for (let i = 0; i < 16; i++) hex += bv[i].toString(16).padStart(2, "0") + " ";
        console.log(`[diag] bytes @ gob: ${hex}`);
        const b2 = resizeAddr.readByteArray(16);
        const bv2 = new Uint8Array(b2);
        let hex2 = "";
        for (let i = 0; i < 16; i++) hex2 += bv2[i].toString(16).padStart(2, "0") + " ";
        console.log(`[diag] bytes @ resize: ${hex2}`);
    } catch (e) { console.log(`[!] byte dump failed: ${e}`); }

    arrayResize = new NativeFunction(resizeAddr, 'int', ['pointer', 'int']);

    // ─── 探针: hook Array::resize — 每帧成千上万次调用, 确认 Frida 能拦截 libgodot_android.so ─
    let resizeHits = 0;
    Interceptor.attach(resizeAddr, {
        onEnter: function(args) { resizeHits++; }
    });
    setTimeout(function() {
        console.log(`[probe] Array::resize 被调用 ${resizeHits} 次 (应该 > 0)`);
    }, 2000);

    // ─── Hook get_overlapping_bodies ─────────────────────────
    let gobCallCount = 0;
    const seenPtrs = new Set();
    Interceptor.attach(gobAddr, {
        onEnter: function(args) {
            // X0 = Area3D *this, X8 = out Array* (hidden struct return)
            this.selfPtr = args[0];
            this.outPtr  = this.context.x8;
            state.lastArea3D = this.selfPtr.toString();
            gobCallCount++;
            const key = this.selfPtr.toString();
            if (!seenPtrs.has(key)) {
                seenPtrs.add(key);
                console.log(`[gob#${gobCallCount}] NEW Area3D this=${key}  outPtr=${this.outPtr}`);
            }
        },
        onLeave: function(retval) {
            // 决定是否要 force:
            //   - 还没发现 Trigger2 (trigger2Ptr == null) → 对所有 Area3D force
            //     一次 (确保 Trigger2 这帧能触发)
            //   - 已经发现 Trigger2 → 只给 Trigger2 这个 this 指针 force
            let shouldForce;
            if (state.trigger2Ptr === null) {
                shouldForce = true;
            } else {
                shouldForce = (this.selfPtr.toString() === state.trigger2Ptr);
            }
            if (!shouldForce) return;

            // 如果原本就有 overlap (真实碰撞), 不动它
            const cur = getArraySize(this.outPtr);
            if (cur > 0) return;

            try {
                arrayResize(this.outPtr, 1);
                state.forceCount++;
                if (state.forceCount <= 5 || state.forceCount % 60 === 0) {
                    console.log(`    [force #${state.forceCount}] this=${this.selfPtr} size 0→${getArraySize(this.outPtr)}`);
                }
            } catch (e) {
                console.log(`[!] resize failed: ${e}`);
            }
        }
    });
    // 每 3 秒打印一次调用计数, 用于诊断 "hook 有没有生效"
    setInterval(function() {
        console.log(`[diag] get_overlapping_bodies 总调用次数 = ${gobCallCount}, force 次数 = ${state.forceCount}, trigger2Ptr = ${state.trigger2Ptr}`);
    }, 3000);
    console.log(`[+] get_overlapping_bodies hooked — 对所有 Area3D force 直到找到 Trigger2`);

    // ─── Hook GameExtension::Process ─────────────────────────
    Interceptor.attach(procAddr, {
        onEnter: function(args) {
            // Process 只在 Trigger2._process 里被调
            // 此时 lastArea3D 就是 Trigger2 的 this 指针
            if (state.trigger2Ptr === null && state.lastArea3D !== null) {
                state.trigger2Ptr = state.lastArea3D;
                console.log(`\n[★] 发现 Trigger2 this ptr = ${state.trigger2Ptr}`);
                console.log(`[★] 从此之后只 force Trigger2, Trigger1 和其他 Area3D 恢复正常`);
            }
            this.retSlot = args[0];
            // 读取 input PackedByteArray (用于 log)
            try {
                const pba = args[2];
                const cp = pba.readPointer();
                if (!cp.isNull()) {
                    const sz = cp.sub(8).readU64().toNumber();
                    if (sz > 0 && sz <= 32) {
                        let hex = "";
                        for (let i = 0; i < sz; i++)
                            hex += cp.add(i).readU8().toString(16).padStart(2, "0");
                        console.log(`    Process input (xor_enc'd token) = ${hex}`);
                    }
                }
            } catch (e) {}
        },
        onLeave: function(retval) {
            // 读 return String
            try {
                const stringPtr = this.retSlot.readPointer();
                if (stringPtr.isNull()) return;
                const length = stringPtr.sub(8).readU64().toNumber();
                if (length <= 0 || length > 64) return;
                let text = "";
                for (let i = 0; i < length; i++) {
                    text += String.fromCodePoint(stringPtr.add(i * 4).readU32());
                }
                state.realFlag = `flag{${FLAG_PREFIX}${text}}`;
                console.log(`\n REAL FLAG = ${state.realFlag}\n`);
                console.log(`force 总次数: ${state.forceCount}`);
            } catch (e) {
                console.log(`[!] 读 Process 返回 String 失败: ${e}`);
            }
        }
    });
    console.log("");
}

installHooks();

这里的token 和flag可以拿到后面验证

静态 SceneTree 还原[赛后]

建议先看后面章节再回来

上面的 Frida 方案虽然能出 flag,但技术味道偏运行时取巧， hook Area3D::get_overlapping_bodies 强制 resize,本质是让游戏自己去执行那段 GDScript 把 flag 写到 Label2。这种做法放到 UE4 里对标,大概等于用 UEDumper 的live 模式打断点抓内存,而不是经典的SDK dumper，照着源码的数据结构,静态枚举对象和读属性

Godot 4 里有没有 UE4 FUObjectArray+ SDK dumper 的对应物?有:

UE4	Godot 4.5
GUObjectArray (FUObjectArray)	ObjectDB::object_slots
UObject::GetName / GetOuter	Node::data.name / .parent
SDK dumper 静态读内存	解 .scn文件 + ObjectDB 内存枚举

但有个更好的做法:Godot 的场景是完整序列化在磁盘上的 .scn(二进制 PackedScene)或 .tscn(文本)。节点类、名字、父子关系、所有属性的初始值 全部写死在文件里。APK 解包后,这些文件就在 export-*.scn根本不需要游戏运行就能完整还原 SceneTree。

decrypt_godot45.py放在了解包章节

decrypt_godot45.py   AES-CFB-128 解密每个 .scn / .gdc
scn_parser.py         Godot 4.5 binary RSRC parser + PackedScene 还原

调用流程:

export-...-town_scene.scn  (AES 加密)
       │
       │  decrypt_godot45.py  (AES-CFB-128 with script_encryption_key)
       ▼
decrypted_scenes/town_scene.scn   (RSRC magic 开头的纯 Godot 资源)
       │
       │  scn_parser.py
       ▼
场景文件

scn 文件是加密的

APK 解包拿到的 .scn首 4 字节不是 RSRC,而是随机字节:

用 decrypt_godot45.py(key 是从 libgodot_android.so .data 段扫 Shannon 熵找到的 32 字节值)解出来

脚本也在解包那里

参考源码，RSRC 格式 (core/io/resource_format_binary.cpp)

Godot 4 的二进制资源文件布局固定:

┌─────────────────────────────────────┐
│ 4B  "RSRC"  magic                    │  header
│ 4B  big_endian                        │
│ 4B  use_real64                        │
│ 4B  engine_ver_major                  │
│ 4B  engine_ver_minor                  │
│ 4B  format_version                    │
│ U*  ustring resource_type             │   "PackedScene"
│ 8B  importmd_ofs                      │
│ 4B  flags                             │   UIDS | NAMED_SCENE_IDS | REAL_T_DOUBLE | HAS_SCRIPT_CLASS
│ 8B  uid          (if UIDS)            │
│ U*  script_class (if HAS_SCRIPT_CLASS)│
│ 44B reserved[11] = 0                  │
├─────────────────────────────────────┤
│ 4B  string_table_size                 │  string table: all names / property keys
│ N * ustring                           │
├─────────────────────────────────────┤
│ 4B  ext_resources_count               │  external refs (scripts, textures, subscenes)
│ N * (type, path, uid?)                │
├─────────────────────────────────────┤
│ 4B  int_resources_count               │  index: (path, byte_offset)
│ N * (path, u64_offset)                │
├─────────────────────────────────────┤
│ 每个 int_resource 在其 offset 处展开: │
│   U*  class_name                      │
│   4B  property_count                  │
│   N * (u32 name_idx, Variant value)   │
├─────────────────────────────────────┤
│ ...                                   │
└─────────────────────────────────────┘

ustring的编码是:u32 length_including_NUL 后跟 length`个 UTF-8 字节(没有 4 字节对齐填充)。

然后Variant 序列化

这里是坑最多的地方。每个 Variant 是 32 type_tag 后跟类型相关字节。Godot 4 的类型枚举:

VARIANT_NIL = 1,  VARIANT_BOOL = 2,  VARIANT_INT = 3,  VARIANT_FLOAT = 4,
VARIANT_STRING = 5,
VARIANT_VECTOR2 = 10, VARIANT_RECT2 = 11, VARIANT_VECTOR3 = 12, VARIANT_PLANE = 13,
VARIANT_QUATERNION = 14, VARIANT_AABB = 15, VARIANT_BASIS = 16,
VARIANT_TRANSFORM3D = 17, VARIANT_TRANSFORM2D = 18, VARIANT_COLOR = 20,
VARIANT_NODE_PATH = 22, VARIANT_OBJECT = 24, VARIANT_DICTIONARY = 26,
VARIANT_ARRAY = 30, VARIANT_PACKED_BYTE_ARRAY = 31, VARIANT_PACKED_INT32_ARRAY = 32,
VARIANT_PACKED_FLOAT32_ARRAY = 33, VARIANT_PACKED_STRING_ARRAY = 34,

VARIANT_PACKED_VECTOR3_ARRAY = 35,
VARIANT_PACKED_COLOR_ARRAY   = 36,
VARIANT_PACKED_VECTOR2_ARRAY = 37,   // 原来是 35
// ..
VARIANT_PACKED_VECTOR4_ARRAY = 53,   // 4.3+ 新增

写 parser 时踩的第一个坑:原本照Godot 4的通用写 PACKED_VECTOR2_ARRAY = 3,解 town_scene 时在某个 CSGPolygon3D 的 polygon 数组处就跑偏了 。实际那个 tag 35 在 4.5 里已经是 PACKED_VECTOR3_ARRAY。

Python 实现:

def _get_string(self) -> str:
    id_ = self.r.u32()
    if id_ & 0x80000000:
        length = id_ & 0x7FFFFFFF
        data = self.r.raw(length)
        if data and data[-1] == 0:
            data = data[:-1]
        return data.decode('utf-8', errors='replace')
    return self.string_map[id_] if id_ < len(self.string_map) else f'<str#{id_}>'

Object 引用

场景里大量字段是引用另一个资源”脚本、mesh、纹理、subscene)。VARIANT_OBJECT的编码:

case VARIANT_OBJECT: {
    uint32_t objtype = f->get_32();
    switch (objtype) {
        case OBJECT_EMPTY: ...                     // 0 = null
        case OBJECT_EXTERNAL_RESOURCE: ...         // 1 = 旧格式 (type+path 字符串)
        case OBJECT_INTERNAL_RESOURCE:             // 2 = 本文件内某个 int_res
            uint32_t index = f->get_32();
        case OBJECT_EXTERNAL_RESOURCE_INDEX:       // 3 = ext_res[] 数组下标
            uint32_t index = f->get_32();
    }
}

在 4.5 文件里几乎只见 2 和 3,都只是一个 u32 索引,很紧凑。

PackedScene 的扁平化节点数组

最核心的一步。解出来的 PackedScene资源只有一个属性 _bundled(Dictionary),里面是完整的场景定义。字段:

_bundled = {
    "names":     PackedStringArray,   # scene-local string pool
    "variants":  Array,               # scene-local variant pool
    "node_count": int,
    "nodes":     PackedInt32Array,    # 节点数据扁平化成 int 数组(核心)
    "conn_count": int,
    "conns":     PackedInt32Array,    # 信号连接
    "node_paths": Array[NodePath],
    "editable_instances": Array[NodePath],
    "version": int,
    "base_scene": int (可选),
}

nodes数组是整棵树的扁平编码。scene/resources/packed_scene.cpp:SceneState::pack() 的写入顺序:

每个节点的连续 int 序列:
  parent_idx      i32     // 父节点在数组里的索引 (-1 = 根)
  owner_idx       i32
  type_idx        i32     // names[type_idx] 是类名;-1/0x3FFFFFFF 表示 instanced scene
  name_idx        i32     // 低 N 位是 names[] 索引
  instance_idx    i32     // variants[instance_idx] 是 subscene 实例(若有)
  property_count  i32
  { name_idx, value_idx } × property_count        // 指向 names[] / variants[]
  group_count     i32
  { name_idx } × group_count

还原算法就一个循环:从头走 nodes[],按这个 schema 解析每个节点。parent_idx 自然形成树。

完整代码

解包

寻找key

没接触过godot，但是应该需要解包，查阅相关资料

DownUnderCTF 2025 - Un1corn - 博客园

查到工具Releases · GDRETools/gdsdecomp

观察结构发现是散包模式资源直接散在assets下，不是打成单个.pck，还有一些so,Godot 导出模板在编译时内置一个 script_encryption_key[32] 全局变量，导出 APK 时编辑器将真实 key 写入该变量位置。key 必然存在于 libgodot_android.so 的 .data段（因为是有显式初始值的可写全局变量）。

根据博客的两种方式定位key寻找失败，似乎是被strip了，但是肯定的是key肯定在libgodot_android里

可以写一个算法专门扫描key，大概思路是

• 前后各8字节全零
• Shannon 熵 ≥ 4.0
• 字节多样性 ≥ 20 种不同值

AES字节的混乱程度为熵

来加快扫描

import math
from collections import Counter

LIB = "preliminary/lib/arm64-v8a/libgodot_android.so"

with open(LIB, "rb") as f:
    blob = f.read()

# Sections we care about (writable + readable + initialized)
SECTIONS = {
    ".rodata":      (0x3aa5d0, 0x625ac8),
    ".data.rel.ro": (0x3ea3470, 0x156370),
    ".data":        (0x3ffdee8, 0x7100),
}

def shannon(b):
    if not b: return 0.0
    c = Counter(b)
    n = len(b)
    return -sum((v/n) * math.log2(v/n) for v in c.values())

def section_for(off):
    for name, (s, sz) in SECTIONS.items():
        if s <= off < s + sz:
            return name
    return "?"

hits = []
for i in range(8, len(blob) - 40):
    pre  = blob[i-8 : i]
    win  = blob[i : i+32]
    post = blob[i+32 : i+40]
    if pre != b"\x00"*8: continue
    if post != b"\x00"*8: continue
    if win.count(0) > 6: continue
    if len(set(win)) < 18: continue
    ent = shannon(win)
    if ent < 4.0: continue
    hits.append((ent, i, win))

hits.sort(reverse=True)
print(f"found {len(hits)} 32B high-entropy islands surrounded by ≥8 bytes of zeros\n")
for ent, off, win in hits[:30]:
    sec = section_for(off)
    print(f"  ent={ent:.2f}  file off=0x{off:08x}  section={sec}")
    print(f"    {win.hex().upper()}")

输出

.data + 0x4f08  (va = 0x400edf0)
CE4DF8753B59A5A39ADE58AC07EF947A3DA39F2AF75E3284D51217C04D49A061

在IDA看

找到key但是用工具解不开，报错Wrong key，猜想可能被魔改AES加密算法?

AES 逆向

问了GPT每个加密的 .gdc/ .gdextension 文件格式，这是Godot常见加密文件头和导出脚本格式(带完整性校验的 AES 加密文件头)

Offset  Size  含义
0x00    16    校验字段（非传统 MD5，可能是 HMAC）
0x10    8     明文大小 (uint64 LE)
0x18    16    AES IV
0x28    N     密文 (N = 明文大小向上对齐到 16 字节)

.gdextension是Godot 4 的原生扩展描述文件。用来告诉 Godot 应该如何加载一个 GDExtension，包括要加载哪些平台下的动态库、入口符号、兼容性信息等；真正执行的是它指向的本地库文件，比如 .dll / .so / .dylib。

.gdc是 GDScript 的编译/加密后的脚本产物。Godot 3.x 里，官方有script encryption key机制，导出时可以把脚本加密，避免以明文形式直接被看到。对应资料里明确提到导出时可对脚本使用 256-bit AES 密钥保护。

追踪调用链：

sub_3804C2C (PCK loader)
  → sub_3801410 (FileAccessEncrypted::open_internal)
    → sub_376EDA0 (mbedtls_aes_init)
    → sub_376EDFC → sub_197C210 (mbedtls_aes_setkey_enc, 256-bit)
    → sub_376EF68 (wrapper)
      → sub_197DE18 (自定义 CFB-128)

关键是识别sub_197C210大量AES的特征

sub_197C210通过 AES key schedule 展开逻辑确认是setkey_enc（正向密钥扩展）

sub_197DE18是自定义的AES 魔改，函数签名如下，与标准 mbedtls CFB-128 的关键差异在于解密分支（mode = 0）：

sub_197DE18(ctx, mode, length, iv_off_ptr, iv, input, output)

标准 mbedtls CFB-128 解密：

output[i] = iv[n] ^ ct[i];
iv[n] = ct[i];
n = (n + 1) & 0xf;

魔改 CFB-128 解密：

output[i] = iv[n] ^ ct[i] ^ n;     //  额外 XOR 块内偏移 n
iv[n] = ct[i] ^ n;                 //  feedback 也 XOR 了 n
n = (n + 1) & 0xf;

Python实现

解压

import os
from cryptography.hazmat.primitives.ciphers import Cipher, algorithms, modes

KEY = bytes.fromhex("CE4DF8753B59A5A39ADE58AC07EF947A3DA39F2AF75E3284D51217C04D49A061")

def aes_enc(key, block):
    """One-block ECB encrypt with AES-256."""
    c = Cipher(algorithms.AES(key), modes.ECB())
    return c.encryptor().update(block)

def godot45_cfb_decrypt(key, iv, ct):
    """Godot 4.5 modified CFB-128 decrypt."""
    iv = bytearray(iv)
    pt = bytearray(len(ct))
    n = 0
    for i in range(len(ct)):
        if n == 0:
            iv = bytearray(aes_enc(key, bytes(iv)))
        c = ct[i]
        pt[i] = iv[n] ^ c ^ n
        iv[n] = c ^ n
        n = (n + 1) & 0xf
    return bytes(pt)

def decrypt_file(path):
    blob = open(path, "rb").read()
    size = int.from_bytes(blob[16:24], "little")
    iv = blob[24:40]
    ct = blob[40:]
    pt = godot45_cfb_decrypt(KEY, iv, ct)
    return pt[:size]

if __name__ == "__main__":
    files = [
        "preliminary/assets/token.gdc",
        "preliminary/assets/label2.gdc",
        "preliminary/assets/spedometer.gdc",
        "preliminary/assets/Trigger/trigger.gdc",
        "preliminary/assets/car_select/car_select.gdc",
        "preliminary/assets/ext/sec2026.gdextension",
    ]
    os.makedirs("decrypted", exist_ok=True)
    for path in files:
        pt = decrypt_file(path)

        out = "decrypted/" + os.path.basename(path)
        with open(out, "wb") as f:
            f.write(pt)

        print(f"=== {path}  size={len(pt)} → {out}")
        if path.endswith(".gdextension"):
            try:
                print(pt.decode("utf-8"))
            except UnicodeDecodeError:
                print("(non-UTF-8)")
                print(pt[:100].hex())
        else:
            for i in range(0, min(64, len(pt)), 16):
                chunk = pt[i:i+16]
                print(f"  +{i:03x}: {chunk.hex()}  {''.join(chr(b) if 32<=b<127 else '.' for b in chunk)}")
        print()

.gdextension解密后是完整的 INI 文本：

反编译

Godot 4.5 gdc 文件格式

Offset  Size  含义
0x00    4     Magic "GDSC"
0x04    4     Bytecode version (uint32 LE, 101 = 0x65 for Godot 4.5)
0x08    4     Uncompressed size (uint32 LE)
0x0C    N     zstd 压缩帧（标识符用 XOR 0xb6 混淆）

使用GDRE Tools反编译,解密后的 .gdc 是标准格式，GDRE Tools 可以直接处理：

./gdre_tools.x86_64 --headless \
    --decompile=decrypted/token.gdc \
    --bytecode=4.5.0 \
    --output=decrypted/decompiled

5 个文件全部成功反编译为 .gd`源码。

可以看到part1

flag生成算法解析

GD分析

flag生成算法为根据屏幕左上角的随机token生成的右上角的flag算法

token.gd

const TOKEN_LEN = 8
const CHARS = "0123456789abcdef"

func generate_token(len: int) -> String:
    var s = ""
    for i in len:
        var idx = rng.randi_range(0, CHARS.length() - 1)
        s += CHARS[idx]
    return s

func _ready():
    rng.randomize()
    text = "Token: " + generate_token(TOKEN_LEN)

Token 是 8 个随机 hex 字符

trigger.gd

const FLAG_PREFIX = "sec2026_PART1_"
var obj = GameExtension.new()

func xor_enc(plain: String) -> PackedByteArray:
    var out_buf = plain.to_utf8_buffer()
    if out_buf.size() < 8: out_buf.resize(8)
    var result = out_buf.slice(0, 8)
    for i in range(7):
        result[i] = result[i] ^ result[i + 1]
    result[7] = result[7] ^ result[0]
    return result

# 当玩家撞到 Trigger2 时触发：
var token_str = str(label1.text).substr(7)      
var flag1 = obj.Process(xor_enc(token_str))     # native 处理
label.text = "flag{" + FLAG_PREFIX + flag1 + "}"

输入[a, b, c, d, e, f, g, h]：

Step 1: for i in 0..6:  result[i] ^= result[i+1]
Step 2: result[7] ^= result[0]  (使用 Step 1 之后的 result[0])

变换后：

[a^b, b^c, c^d, d^e, e^f, f^g, g^h, h^(a^b)]

xor_enc 逆运算

给定输出 [x0, x1, x2, x3, x4, x5, x6, x7]，还原原始 [a, b, c, d, e, f, g, h]：

def xor_enc_inverse(x):
    h = x[7] ^ x[0]      # x7 = h ^ x0，所以 h = x7 ^ x0
    g = x[6] ^ h
    f = x[5] ^ g
    e = x[4] ^ f
    d = x[3] ^ e
    c = x[2] ^ d
    b = x[1] ^ c
    a = x[0] ^ b
    return bytes([a, b, c, d, e, f, g, h])

flag1还要加上Process的处理

Dump 真实so文件

发现libsec2026.so没几个函数，字符串没有交叉引用

nm libsec2026.so ，T extension_init 0x56d50

readelf 0x56d50 在 LOAD 段（perm=R+X，应该是代码）

但 get_bytes(0x56d50) 读出来完全不像 ARM64 指令

应该是被壳或者其它东西保护了

start 是合法的 ARM64 代码，是加壳器的解密 stub，运行时把真正的代码解密到内存中。

; 解析 /proc/self/auxv 获取页大小
svc #0 (openat "/proc/self/auxv")
svc #0 (read)
svc #0 (close)

; 调用解压器
BL  sub_69984                      ; sub_69984(&dword_69A6C, 4129, out_buf, &out_size)

; 创建匿名 memfd
svc #0 (memfd_create)
svc #0 (write)
svc #0 (mmap PROT_READ|PROT_EXEC)
svc #0 (close)

; 跳转到解压后的代码 + 0x10
ADR x0, qword_69860
ADD x14, x14, #0x10
BR  x14                             ; jump to unpacked code (skipping 16-byte header)

Header 在 0x69A60：

偏移	字段
0x69A60	未压缩大小 0x17A0
0x69A6	压缩大小 0x1021
0x69A68	unknown (= 2)
0x69A6C	压缩数据起点

这样就有两种方法了,静态逆向，动态dump

方法1: 静态逆向

分析剩下的代码发现实现了以下功能

sub_69984（64 字节）+ sub_699C4（24 字节位读取器）+ sub_699DC（gamma 码读取器）

是aPlib变种，其中特征

• 32-bit 位缓冲 (adds w4, w4, w4 + ldr w4, [x0], #4 )
• 每符号 1 bit 的 literal/match 标志
• Gamma 码变长 length 编码（交替读 data bit 和 control bit）
• cmn w5, #0xD00 长距离优化阈值
• 终止符 = 距离 = 0xFFFFFFFF (mvn w5, w5; cbz w5)

让AI搓了一个脚本

import struct

def decode_aplib(src: bytes) -> bytes:
    src_pos = [0]
    bitbuf = [0x80000000]

    def read_bit():
        val = bitbuf[0]
        new_val = (val << 1) & 0xFFFFFFFF
        old_carry = (val >> 31) & 1 
        if new_val == 0:
            dword = int.from_bytes(src[src_pos[0]:src_pos[0] + 4], "little")
            src_pos[0] += 4
            bitbuf[0] = ((dword << 1) + old_carry) & 0xFFFFFFFF
            return (dword >> 31) & 1
        bitbuf[0] = new_val
        return old_carry

    def read_gamma():
        val = 1
        while True:
            val = (val << 1) | read_bit()
            ctrl = read_bit()
            if ctrl == 1:
                return val

    def read_lit():
        b = src[src_pos[0]]
        src_pos[0] += 1
        return b

    dst = bytearray()
    distance_neg = 0xFFFFFFFF 

    while True:
        if read_bit() == 1:
            # Literal byte
            dst.append(read_lit())
            continue

        # Match
        gamma = read_gamma()
        if gamma >= 3:
            byte = read_lit()
            new_distance_neg = ((gamma - 3) << 8) | byte
            distance_neg = (~new_distance_neg) & 0xFFFFFFFF
            if distance_neg == 0: 
                break

        w1 = read_bit()
        w1 = (w1 << 1) | read_bit()
        if w1 == 0:
            w1 = read_gamma() + 2

        distance = (0x100000000 - distance_neg) & 0xFFFFFFFF
        if distance > 0xD00:
            w1 += 1

        length = w1 + 1
        for _ in range(length):
            dst.append(dst[len(dst) - distance])

    return bytes(dst)

if __name__ == "__main__":
    import os
    import sys

    LIB_PATH = "preliminary/lib/arm64-v8a/libsec2026.so"
    BLOB_OFFSET = 0x69A6C  # from sub_69984 call in `start`
    BLOB_SIZE = 4129        # from header at 0x69A60+4
    EXPECTED_SIZE = 6048    # from header at 0x69A60+0

    if not os.path.exists(LIB_PATH):
        print(f"错误: 找不到 {LIB_PATH}", file=sys.stderr)
        print("请在项目根目录 (preliminary 的父目录) 运行此脚本", file=sys.stderr)
        sys.exit(1)

    with open(LIB_PATH, "rb") as f:
        lib_bytes = f.read()

    # Also verify the header at 0x69A60
    header = lib_bytes[0x69A60:0x69A6C]
    usize, csize, flag = struct.unpack("<III", header)
    print(f"Stage1 header @ 0x69A60:")
    print(f"  uncompressed size = 0x{usize:x} = {usize}")
    print(f"  compressed size   = 0x{csize:x} = {csize}")
    print(f"  unknown field     = 0x{flag:x}")
    assert usize == EXPECTED_SIZE, f"unexpected usize {usize}"
    assert csize == BLOB_SIZE, f"unexpected csize {csize}"

    blob = lib_bytes[BLOB_OFFSET:BLOB_OFFSET + csize]
    print(f"\n解压 {len(blob)} 字节 aPlib 压缩数据...")
    out = decode_aplib(blob)
    print(f"解压完成: {len(out)} 字节 (期望 {EXPECTED_SIZE})")
    assert len(out) == EXPECTED_SIZE

    out_path = "unpacked_stage2.bin"
    with open(out_path, "wb") as f:
        f.write(out)
    print(f"\n已保存到 {out_path}")
    print(f"\n前 16 字节 (header):     {out[:16].hex()}")
    print(f"后接 4 字节 (entry):     {out[16:20].hex()}  (= {int.from_bytes(out[16:20], 'little'):08x})")

坑点：gamma 码的 control bit 语义，b.cc loop意味着 control=0 继续循环，control=1停止，第一次写反了导致完全解错

把 6048 字节的 stage 2 加载到新 IDA segment 分析

#!/usr/bin/env python3
import ida_bytes
import ida_segment
import ida_ua
import ida_funcs
import ida_auto
import idaapi
import idautils
import idc

def decode_aplib(src: bytes) -> bytes:
    src_pos = [0]
    bitbuf = [0x80000000]

    def read_bit():
        val = bitbuf[0]
        new_val = (val << 1) & 0xFFFFFFFF
        old_carry = (val >> 31) & 1
        if new_val == 0:
            dword = int.from_bytes(src[src_pos[0]:src_pos[0] + 4], "little")
            src_pos[0] += 4
            bitbuf[0] = ((dword << 1) + old_carry) & 0xFFFFFFFF
            return (dword >> 31) & 1
        bitbuf[0] = new_val
        return old_carry

    def read_gamma():
        val = 1
        while True:
            val = (val << 1) | read_bit()
            if read_bit() == 1:
                return val

    def read_lit():
        b = src[src_pos[0]]
        src_pos[0] += 1
        return b

    dst = bytearray()
    distance_neg = 0xFFFFFFFF
    while True:
        if read_bit() == 1:
            dst.append(read_lit())
            continue
        gamma = read_gamma()
        if gamma >= 3:
            byte = read_lit()
            new_neg = ((gamma - 3) << 8) | byte
            distance_neg = (~new_neg) & 0xFFFFFFFF
            if distance_neg == 0:
                break
        w1 = read_bit()
        w1 = (w1 << 1) | read_bit()
        if w1 == 0:
            w1 = read_gamma() + 2
        distance = (0x100000000 - distance_neg) & 0xFFFFFFFF
        if distance > 0xD00:
            w1 += 1
        for _ in range(w1 + 1):
            dst.append(dst[len(dst) - distance])
    return bytes(dst)


# libsec2026.so 里的 aPlib blob 位置
BLOB_VADDR = 0x69A6C
BLOB_SIZE = 4129

# 在 IDA 里选一个远离现有 segments 的高地址映射 stage 2
SEG_BASE = 0x200000
SEG_SIZE = 0x2000           # 6048 字节 < 0x2000, 足够
ENTRY_VADDR = SEG_BASE + 0x10  # 真正的 entry = offset 0x10 (跳过 16B header)

print(f"[*] 从 libsec2026.so @ {hex(BLOB_VADDR)} 读取 {BLOB_SIZE} 字节 aPlib 压缩数据")
compressed = ida_bytes.get_bytes(BLOB_VADDR, BLOB_SIZE)
if compressed is None or len(compressed) != BLOB_SIZE:
    raise RuntimeError(f"无法读取 blob from {hex(BLOB_VADDR)}, 确认当前 session 是 libsec2026.so")

print(f"[*] 解压 aPlib...")
unpacked = decode_aplib(compressed)
print(f"[+] 解压得到 {len(unpacked)} 字节")
assert len(unpacked) == 6048, f"期望 6048 字节, 实际 {len(unpacked)}"

# ─── 创建 segment ───
existing = ida_segment.get_segm_by_name("unpacked")
if existing:
    print(f"[!] segment 'unpacked' 已存在, 跳过创建")
else:
    print(f"[*] 创建 segment 'unpacked' @ {hex(SEG_BASE)}..{hex(SEG_BASE + SEG_SIZE)}")
    ok = idaapi.add_segm(
        0,                       # selector
        SEG_BASE,                # start_ea
        SEG_BASE + SEG_SIZE,     # end_ea
        "unpacked",              # name
        "CODE",                  # class: CODE means executable
    )
    if not ok:
        raise RuntimeError("add_segm 失败")

    # 设置 64-bit 模式 + R+X 权限
    seg = ida_segment.get_segm_by_name("unpacked")
    seg.bitness = 2  # 2 = 64-bit
    seg.perm = 5      # 5 = READ | EXEC
    ida_segment.update_segm(seg)

# ─── 写入解压后的字节 ───
print(f"[*] 写入 {len(unpacked)} 字节到 {hex(SEG_BASE)}")
for i, b in enumerate(unpacked):
    ida_bytes.patch_byte(SEG_BASE + i, b)

# ─── 强制 entry point 分析为代码 ───
print(f"[*] 在 {hex(ENTRY_VADDR)} (stage 2 entry) 创建函数")
# 先 undefine 整段, 避免 IDA 把它当作 data
ida_bytes.del_items(ENTRY_VADDR, ida_bytes.DELIT_SIMPLE, SEG_SIZE - 0x10)
# 让 IDA 从 entry 开始分析
ida_ua.create_insn(ENTRY_VADDR)
ida_funcs.add_func(ENTRY_VADDR)

# ─── 触发自动分析覆盖整个 segment ───
print(f"[*] 触发自动分析 {hex(SEG_BASE)}..{hex(SEG_BASE + SEG_SIZE)}")
ida_auto.plan_and_wait(SEG_BASE, SEG_BASE + SEG_SIZE)

# ─── 打印识别出的函数 ───
funcs = [f for f in idautils.Functions(SEG_BASE, SEG_BASE + SEG_SIZE)]
print(f"\n[+] 自动分析完成! 在 stage 2 里识别出 {len(funcs)} 个函数:")
for fn_ea in funcs[:30]:
    fn = ida_funcs.get_func(fn_ea)
    name = ida_funcs.get_func_name(fn_ea)
    size = fn.end_ea - fn.start_ea
    print(f"  {hex(fn_ea)}  {name:<24}  size={hex(size)}")
if len(funcs) > 30:
    print(f"  ... 还有 {len(funcs) - 30} 个函数未显示")

执行效果

在sub_201150发现

0x2158055=UPX!，是个UPX壳

12 字节 chunk header 格式：

偏移	字段
+0	orig_size (uint32) — 解压后大小
+4	comp_size (uint32) — 压缩数据大小
+8	method (uint8) — 2=aPlib, 14=LZMA
+9	filter (uint8) — 0 或 82='R' (ARM64 BL 重定位修正)
+10..12	填充

LZMA 魔改

sub_200394 是 LZMA range decoder，也是魔改：

参数	标准 LZMA	本题变种
kBitModelTotal	2048 (11-bit)	4096 (12-bit)
初始 prob	1024	2048
范围检查	range >> 11	range >> 12
kNumMoveBit	5	5

而且属性字节 (lc, lp, pb) 编码方式也不同（3 个分开的字节而不是打包的 (pb*5+lp)*9+lc）：

// sub_20024C 里的 LZMA 方法 14 初始化：
pb = byte0 & 7           // 位置比特
lc = byte1 >> 4          // 字面量上下文比特
lp = byte1 & 0xF         // 字面量位置比特

Python标准lzma模块无法解码这个变种

BL 重定位过滤器（filter=82 ‘R’），LZMA 解压后还有一层 AArch64 BL/B 指令修正：

for (i = num_insns - 1; i >= 0; i--) {
    insn = *(uint32_t*)(dst + 4*i);
    if (((insn >> 26) & 0x1F) == 5) {
        *(uint32_t*)(dst + 4*i) = (insn & 0xFC000000) | ((insn - i) & 0x3FFFFFF);
    }
}

把绝对偏移形式的 BL 转回相对偏移形式

因为 LZMA 变种太复杂，手写 Python 移植太耗时，用 Unicorn 直接模拟执行 sub_20024C：

import struct
import os
import sys
from unicorn import Uc, UC_ARCH_ARM64, UC_MODE_ARM, UC_PROT_ALL
from unicorn.arm64_const import (
    UC_ARM64_REG_X0, UC_ARM64_REG_X1, UC_ARM64_REG_X2, UC_ARM64_REG_X3,
    UC_ARM64_REG_W1, UC_ARM64_REG_W4, UC_ARM64_REG_SP, UC_ARM64_REG_LR,
    UC_ARM64_REG_PC,
)

LIBSEC_PATH = "preliminary/lib/arm64-v8a/libsec2026.so"
STAGE2_PATH = "unpacked_stage2.bin"

# Stage 2 代码在 Unicorn 里的映射位置
STAGE2_BASE = 0x200000
STAGE2_SIZE = 0x2000

SUB_20024C_OFFSET = 0x24C   # LZMA/aPlib 解压分发器
SUB_200074_OFFSET = 0x074   

CHUNK2_FILE_OFFSET_IN_LIB = 0x4b147       
CHUNK2_HEADER_SIZE = 12
CHUNK2_ORIG_SIZE = 637936  
CHUNK2_COMP_SIZE = 124681  
CHUNK2_METHOD = 14          
CHUNK2_FILTER = 82         


if not os.path.exists(STAGE2_PATH):
    sys.exit(1)

if not os.path.exists(LIBSEC_PATH):
    print(f"错误: 找不到 {LIBSEC_PATH}", file=sys.stderr)
    sys.exit(1)

stage2 = open(STAGE2_PATH, "rb").read()
assert len(stage2) == 6048, f"stage2 size is {len(stage2)}, expected 6048"

lib = open(LIBSEC_PATH, "rb").read()

chunk2_data = lib[CHUNK2_FILE_OFFSET_IN_LIB + CHUNK2_HEADER_SIZE
                  : CHUNK2_FILE_OFFSET_IN_LIB + CHUNK2_HEADER_SIZE + CHUNK2_COMP_SIZE]
assert len(chunk2_data) == CHUNK2_COMP_SIZE
print(f"[*] chunk 2 压缩数据: {len(chunk2_data)} 字节, 前 8 字节 = {chunk2_data[:8].hex()}")
print(f"    (其中前 2 字节 '1a 03' 是 LZMA props: pb=2, lc=3, lp=0)")


print("\n[*] 初始化 Unicorn (ARM64)")
uc = Uc(UC_ARCH_ARM64, UC_MODE_ARM)

# 映射 stage 2 代码
uc.mem_map(STAGE2_BASE, STAGE2_SIZE, UC_PROT_ALL)
uc.mem_write(STAGE2_BASE, stage2)
print(f"    stage 2 代码 @ {hex(STAGE2_BASE)}..{hex(STAGE2_BASE + STAGE2_SIZE)}")

# 映射栈
STACK_BASE = 0x400000
STACK_SIZE = 0x100000
uc.mem_map(STACK_BASE, STACK_SIZE, UC_PROT_ALL)
print(f"    栈         @ {hex(STACK_BASE)}..{hex(STACK_BASE + STACK_SIZE)}")

# 映射 chunk 2 压缩输入
INPUT_BASE = 0x500000
INPUT_SIZE = 0x40000
uc.mem_map(INPUT_BASE, INPUT_SIZE, UC_PROT_ALL)
uc.mem_write(INPUT_BASE, chunk2_data)
print(f"    压缩输入   @ {hex(INPUT_BASE)} ({len(chunk2_data)} 字节)")

# 映射解压输出缓冲区 (需要至少 637936 字节, 分配 1 MB)
OUTPUT_BASE = 0x600000
OUTPUT_SIZE = 0x100000
uc.mem_map(OUTPUT_BASE, OUTPUT_SIZE, UC_PROT_ALL)
print(f"    解压输出   @ {hex(OUTPUT_BASE)}..{hex(OUTPUT_BASE + OUTPUT_SIZE)}")

# 映射 scratch 区: 放一个 uint32 size_ptr (输出: 实际解压了多少字节)
SCRATCH_BASE = 0x900000
uc.mem_map(SCRATCH_BASE, 0x10000, UC_PROT_ALL)
# 预先写入期望的 orig_size (sub_20024C 的一些路径会读这个值)
uc.mem_write(SCRATCH_BASE, struct.pack("<I", CHUNK2_ORIG_SIZE))
print(f"    size_ptr   @ {hex(SCRATCH_BASE)} (预置 = {CHUNK2_ORIG_SIZE})")

print(f"\n[*] 调用 sub_20024C @ {hex(STAGE2_BASE + SUB_20024C_OFFSET)}")
print(f"    X0 (input ptr)  = {hex(INPUT_BASE)}")
print(f"    W1 (comp_size)  = {CHUNK2_COMP_SIZE}")
print(f"    X2 (output ptr) = {hex(OUTPUT_BASE)}")
print(f"    X3 (size ptr)   = {hex(SCRATCH_BASE)}")
print(f"    W4 (method)     = {CHUNK2_METHOD} (LZMA)")

uc.reg_write(UC_ARM64_REG_X0, INPUT_BASE)
uc.reg_write(UC_ARM64_REG_W1, CHUNK2_COMP_SIZE)
uc.reg_write(UC_ARM64_REG_X2, OUTPUT_BASE)
uc.reg_write(UC_ARM64_REG_X3, SCRATCH_BASE)
uc.reg_write(UC_ARM64_REG_W4, CHUNK2_METHOD)
uc.reg_write(UC_ARM64_REG_SP, STACK_BASE + STACK_SIZE - 0x10000)
uc.reg_write(UC_ARM64_REG_LR, 0xDEADBEEF)

try:
    uc.emu_start(STAGE2_BASE + SUB_20024C_OFFSET, 0xDEADBEEF, timeout=60_000_000)
    print("[+] 模拟执行完成")
except Exception as e:
    pc = uc.reg_read(UC_ARM64_REG_PC)
    print(f"[!] 模拟异常 @ pc={hex(pc)}: {e}")
    sys.exit(1)


raw_output = bytes(uc.mem_read(OUTPUT_BASE, CHUNK2_ORIG_SIZE))
print(f"\n[+] LZMA 解压完成, raw 输出: {len(raw_output)} 字节")
print(f"    前 64 字节: {raw_output[:64].hex()}")

# 应用 BL 重定位 filter
print(f"\n[*] 应用 BL 重定位 filter (filter={CHUNK2_FILTER}=R)")
filtered = bytearray(raw_output)
num_insns = len(filtered) // 4
for i in range(num_insns - 1, -1, -1):
    insn = struct.unpack_from("<I", filtered, i * 4)[0]
    if ((insn >> 26) & 0x1F) == 5:
        new_insn = (insn & 0xFC000000) | ((insn - i) & 0x3FFFFFF)
        struct.pack_into("<I", filtered, i * 4, new_insn)
filtered = bytes(filtered)
print(f"[+] Filter 完成, 最终字节: {filtered[:64].hex()}")

# 用 capstone 反汇编前几条指令
try:
    import capstone
    cs = capstone.Cs(capstone.CS_ARCH_ARM64, capstone.CS_MODE_ARM)
    print("\n[*] 前 12 条 ARM64 指令 (filter 后):")
    for insn in list(cs.disasm(filtered[:48], 0))[:12]:
        print(f"    +{insn.address:04x}: {insn.mnemonic} {insn.op_str}")
except ImportError:
    print("")

# 保存结果
open("chunk2_raw.bin", "wb").write(raw_output)
open("chunk2_filtered.bin", "wb").write(filtered)
print("\n[+] 已保存:")
print(f"    chunk2_raw.bin      ({len(raw_output)} 字节, 未应用 filter)")
print(f"    chunk2_filtered.bin ({len(filtered)} 字节, 应用 BL filter)")
print("\n[*] 这 637,936 字节会被 solve.md 6.8 里的重建脚本放到")
print("    libsec2026_v3.so 的文件偏移 0x4b070..0xe6c60 (= PT_LOAD[1] 主代码段)")

py_eval 环境下 hook 函数的闭包不能访问外部变量，要用 builtins._emu_state 之类的 hack 传状态

需要在IDApython的环境先安装unicorn

输出

first 64 of output:
5f2403d5 1f2003d5 405f4e10 0c6e0214
5f2403d5 c0035fd6 5f2403d5 8e5c0214
5f2403d5 600000b4 f00300aa 00021fd6
c0035fd6 5f2403d5 e10300aa 1f2003d5

可以当AArch64 指令解出来解释处理，说明 LZMA 解压成功。但这时还没做 BL 重定位过滤器 (filter=82=’R’)，所以里面的 B/BL 指令的偏移还是存储优化后的值，需要再跑一遍 filter 修正

加密区布局与 chunks 位置

加密区在 0x4b070 - 0x69860（125424 字节）

外层 24 字节 header 0x4b070：

偏移	值	含义
+0..4	2f ef 94 51	magic
+4..8	20 20 20 20	填充
+8..16	d0 12 0e 2a 00 00 00 00
+16..20	b0 c0 0e 00 = 0xec0b0	总未压缩大小 = 966832
+20..24	60 6c 0e 00 = 0xe6c60	PT_LOAD[1] 大小

Chunk 列表（从 0x4b088 开始）：

#	位置	orig	comp	method	filter	内容
1	+0x18	568	179	2 (aPlib)	0	ELF header + 9 个 phdrs
2	+0xd7	637,936	124,681	14 (LZMA)	82 (‘R’)	主代码段的一部分

只有 2 个 chunks，总计 638504 字节解压输出，远小于966832字节总大小

分析 sub_201450在第一次 sub_201150 之后的 phdr 循环发现：

201578: ldr x20, [x19, #0x10]  ; x20 = phdr.p_vaddr
20157c: ldr x0, [x19, #0x20]   ; x0 = phdr.p_filesz
201580: add w20, w20, w0        ; x20 = vaddr + filesz (segment end)
...
2015b8: sub x1, x20, x1         ; x1 = segment_end - orig_size  ← 放置地址
2015c4: str x1, [x29, #0xd8]    ; var_18 = output destination

**即每个 chunk 的输出放置在 (phdr.p_vaddr + phdr.p_filesz - chunk.orig_size)**，也就是从该 segment 的末尾向前对齐。

对PT_LOAD[1]（vaddr=0, filesz=0xe6c60），chunk 2 (637936 bytes) 的放置地址：

var_18 = 0 + 0xe6c60 - 637936 = 0xe6c60 - 0x9bbf0 = 0x4b070

所以 chunk 2 的输出必须写到文件偏移 0x4b070，而不是 chunk 1 后面的 568

需要重建解压出来的这块数据，因为没有 ELF header，也没有 PHDRs，IDA 加载不了这些数据

第一次重建的策略是把PACKED的vaddr内容直接复制到 REAL 的同 file offset，这样字符串搜索（用 recovered.find(b”Process”)）确实能找到Process在文件里的某个偏移，看起来像成功了。但是把libsec2026_v2.so加载到 IDA发现：

get_bytes(0xed5c4, 20) → '\x60\xe2\x0c\x00...'  # 不是 "Process"，是指针

IDA 按 REAL PHDRs 的 file-vaddr 映射去读 vaddr 0xed5c4，读到的是文件偏移 0xeb5c4 (因为 PT_LOAD[3] file 0xeb5c0 → vaddr 0xed5c0)，但 v2 放到文件偏移 0xeb5c4 的是 .data.rel.ro 的指针内容，不是 .data 的Process字符串。

两份 libsec2026.so 的phdrs 完全不一样，因为 packer 把原 ELF 完全重排了

• PACKED 的 .data (Process) 字符串: vaddr 0xed5c4 = file 0x705c4
• REAL 的 .data (Process”) 字符串: vaddr 0xed5c4 = file 0xeb5c4

在v2里把 PACKED 的 vaddr 0xed5c4 内容放到了 REAL file 0xed5c4，但 IDA 加载时按 REAL PHDRs 去读 REAL file 0xeb5c4，读到错的内容

正确的做法按 REAL PHDRs 的 file↔vaddr 映射把 PACKED vaddr 读到的内容放到 REAL file offset，让AI搓了一份出来

import os
import struct
import sys

# 1. aPlib 解码

def decode_aplib(src: bytes) -> bytes:
    """
    aPlib 变种解码:
      - 32-bit 位缓冲器 (sub_699C4)
      - Gamma 码变长 length (sub_699DC, control bit == 1 停止)
      - 终止符 = 距离 = 0xFFFFFFFF
    """
    src_pos = [0]
    bitbuf = [0x80000000]

    def read_bit():
        val = bitbuf[0]
        new_val = (val << 1) & 0xFFFFFFFF
        old_carry = (val >> 31) & 1
        if new_val == 0:
            dword = int.from_bytes(src[src_pos[0]:src_pos[0] + 4], "little")
            src_pos[0] += 4
            bitbuf[0] = ((dword << 1) + old_carry) & 0xFFFFFFFF
            return (dword >> 31) & 1
        bitbuf[0] = new_val
        return old_carry

    def read_gamma():
        val = 1
        while True:
            val = (val << 1) | read_bit()
            if read_bit() == 1:
                return val

    def read_lit():
        b = src[src_pos[0]]
        src_pos[0] += 1
        return b

    dst = bytearray()
    distance_neg = 0xFFFFFFFF
    while True:
        if read_bit() == 1:
            dst.append(read_lit())
            continue
        gamma = read_gamma()
        if gamma >= 3:
            byte = read_lit()
            new_neg = ((gamma - 3) << 8) | byte
            distance_neg = (~new_neg) & 0xFFFFFFFF
            if distance_neg == 0:
                break
        w1 = read_bit()
        w1 = (w1 << 1) | read_bit()
        if w1 == 0:
            w1 = read_gamma() + 2
        distance = (0x100000000 - distance_neg) & 0xFFFFFFFF
        if distance > 0xD00:
            w1 += 1
        for _ in range(w1 + 1):
            dst.append(dst[len(dst) - distance])
    return bytes(dst)



# 2. Unicorn 模拟 stage 2 的 LZMA 解压器
def unicorn_lzma_decompress(stage2_code: bytes, compressed: bytes, orig_size: int) -> bytes:
    """
    加载 stage 2 loader 到 Unicorn, 调 sub_20024C(method=14) 解压 LZMA。

    sub_20024C 签名:
        (a1=ptr_to_compressed, w1=comp_size, a2=output_buf, a3=size_ptr, w4=method)
    """
    try:
        from unicorn import Uc, UC_ARCH_ARM64, UC_MODE_ARM, UC_PROT_ALL
        from unicorn.arm64_const import (
            UC_ARM64_REG_X0, UC_ARM64_REG_W1, UC_ARM64_REG_X2, UC_ARM64_REG_X3,
            UC_ARM64_REG_W4, UC_ARM64_REG_SP, UC_ARM64_REG_LR,
        )
    except ImportError:
        print("错误: 需要 unicorn 模块。运行: pip install unicorn", file=sys.stderr)
        sys.exit(1)

    STAGE2_BASE = 0x200000
    STACK_BASE  = 0x400000
    INPUT_BASE  = 0x500000
    OUTPUT_BASE = 0x600000
    SCRATCH     = 0x900000
    SUB_20024C_OFFSET = 0x24C

    uc = Uc(UC_ARCH_ARM64, UC_MODE_ARM)
    uc.mem_map(STAGE2_BASE, 0x2000, UC_PROT_ALL)
    uc.mem_write(STAGE2_BASE, stage2_code)
    uc.mem_map(STACK_BASE, 0x100000, UC_PROT_ALL)
    uc.mem_map(INPUT_BASE, 0x40000, UC_PROT_ALL)
    uc.mem_write(INPUT_BASE, compressed)
    uc.mem_map(OUTPUT_BASE, 0x100000, UC_PROT_ALL)
    uc.mem_map(SCRATCH, 0x1000, UC_PROT_ALL)
    uc.mem_write(SCRATCH, struct.pack("<I", orig_size))

    uc.reg_write(UC_ARM64_REG_X0, INPUT_BASE)
    uc.reg_write(UC_ARM64_REG_W1, len(compressed))
    uc.reg_write(UC_ARM64_REG_X2, OUTPUT_BASE)
    uc.reg_write(UC_ARM64_REG_X3, SCRATCH)
    uc.reg_write(UC_ARM64_REG_W4, 14)  # LZMA
    uc.reg_write(UC_ARM64_REG_SP, STACK_BASE + 0x90000)
    uc.reg_write(UC_ARM64_REG_LR, 0xDEADBEEF)

    uc.emu_start(STAGE2_BASE + SUB_20024C_OFFSET, 0xDEADBEEF, timeout=60_000_000)
    return bytes(uc.mem_read(OUTPUT_BASE, orig_size))


def apply_bl_filter(data: bytes) -> bytes:
    """
    AArch64 BL 重定位修正 filter (filter byte = 82 = 'R'):
    对应 sub_200148 里的后处理:
        for i in range(len_insns - 1, -1, -1):
            if ((insn >> 26) & 0x1F) == 5:   # B or BL opcode
                insn = (insn & 0xFC000000) | ((insn - i) & 0x3FFFFFF)
    """
    filtered = bytearray(data)
    num_insns = len(filtered) // 4
    for i in range(num_insns - 1, -1, -1):
        insn = struct.unpack_from("<I", filtered, i * 4)[0]
        if ((insn >> 26) & 0x1F) == 5:
            new_insn = (insn & 0xFC000000) | ((insn - i) & 0x3FFFFFF)
            struct.pack_into("<I", filtered, i * 4, new_insn)
    return bytes(filtered)


# 3. PACKED ELF phdrs 辅助

def parse_elf_phdrs(elf_bytes: bytes):
    e_phoff = struct.unpack("<Q", elf_bytes[32:40])[0]
    e_phnum = struct.unpack("<H", elf_bytes[56:58])[0]
    phdrs = []
    for i in range(e_phnum):
        po = e_phoff + i * 56
        p_type   = struct.unpack("<I", elf_bytes[po:po+4])[0]
        p_offset = struct.unpack("<Q", elf_bytes[po+8:po+16])[0]
        p_vaddr  = struct.unpack("<Q", elf_bytes[po+16:po+24])[0]
        p_filesz = struct.unpack("<Q", elf_bytes[po+32:po+40])[0]
        phdrs.append((p_type, p_offset, p_vaddr, p_filesz))
    return phdrs


def read_at_vaddr(elf_bytes: bytes, phdrs, vaddr: int, size: int) -> bytes:
    """按 phdrs 的 vaddr→file 映射从 elf_bytes 里读 size 字节。"""
    result = bytearray(size)
    for p_type, p_off, p_va, p_fs in phdrs:
        if p_type != 1:  # PT_LOAD only
            continue
        if vaddr + size <= p_va or vaddr >= p_va + p_fs:
            continue
        start_v = max(vaddr, p_va)
        end_v   = min(vaddr + size, p_va + p_fs)
        for v in range(start_v, end_v):
            result[v - vaddr] = elf_bytes[p_off + (v - p_va)]
    return bytes(result)


# 4. 主流程

def main():
    PACKED_PATH = "preliminary/lib/arm64-v8a/libsec2026.so"
    OUTPUT_PATH = "libsec2026_v3.so"

    # 最外层 aPlib blob — 生成 stage 2 loader
    STAGE1_BLOB_VADDR = 0x69A6C
    STAGE1_BLOB_SIZE  = 4129
    STAGE2_EXPECTED   = 6048

    # 加密代码区布局
    ENC_BASE       = 0x4b070
    CHUNK1_DATA    = 0x4b094    # chunk 1 压缩数据位置 (179 B aPlib)
    CHUNK1_ORIG    = 568
    CHUNK1_COMP    = 179
    CHUNK2_DATA    = 0x4b153    # chunk 2 压缩数据位置 (124681 B LZMA)
    CHUNK2_ORIG    = 637936
    CHUNK2_COMP    = 124681

    TOTAL_SIZE     = 966832     # 0xec0b0, 来自 outer header

    if not os.path.exists(PACKED_PATH):
        print(f"错误: 找不到 {PACKED_PATH}", file=sys.stderr)
        print("请在 preliminary/ 的父目录运行此脚本", file=sys.stderr)
        sys.exit(1)

    print(f"[*] 读取 packed: {PACKED_PATH}")
    packed = open(PACKED_PATH, "rb").read()
    print(f"    大小: {len(packed)} 字节")

    # ─── Step 1: 解压 Stage 2 loader ──────────────────────────────────
    print(f"\n[*] Step 1: aPlib 解压 stage 2 loader ({STAGE1_BLOB_SIZE}B → {STAGE2_EXPECTED}B)")
    stage1_blob = packed[STAGE1_BLOB_VADDR:STAGE1_BLOB_VADDR + STAGE1_BLOB_SIZE]
    stage2_loader = decode_aplib(stage1_blob)
    assert len(stage2_loader) == STAGE2_EXPECTED, \
        f"stage2 size {len(stage2_loader)} != {STAGE2_EXPECTED}"
    print(f"[+] 得到 stage 2 loader: {len(stage2_loader)} 字节")

    # ─── Step 2: 解压 Chunk 1 ─────────────────────────────────────────
    print(f"\n[*] Step 2: aPlib 解压 chunk 1 ({CHUNK1_COMP}B → {CHUNK1_ORIG}B, ELF header+phdrs)")
    chunk1_blob = packed[CHUNK1_DATA:CHUNK1_DATA + CHUNK1_COMP]
    chunk1_output = decode_aplib(chunk1_blob)
    assert len(chunk1_output) == CHUNK1_ORIG, \
        f"chunk1 size {len(chunk1_output)} != {CHUNK1_ORIG}"
    assert chunk1_output[:4] == b"\x7fELF", \
        f"chunk1 不是 ELF magic: {chunk1_output[:4].hex()}"
    print(f"[+] ELF magic ✓, phnum = {struct.unpack('<H', chunk1_output[56:58])[0]}")

    # ─── Step 3: 解压 Chunk 2 (LZMA via Unicorn, 然后 BL filter) ──────
    print(f"\n[*] Step 3: Unicorn 模拟 LZMA 解压 chunk 2 ({CHUNK2_COMP}B → {CHUNK2_ORIG}B)")
    chunk2_comp = packed[CHUNK2_DATA:CHUNK2_DATA + CHUNK2_COMP]
    chunk2_raw = unicorn_lzma_decompress(stage2_loader, chunk2_comp, CHUNK2_ORIG)
    assert len(chunk2_raw) == CHUNK2_ORIG
    print(f"[+] LZMA 解压 ✓, 应用 BL 重定位 filter")
    chunk2_filtered = apply_bl_filter(chunk2_raw)

    # ─── Step 4: 解析 PACKED phdrs ───────────────────────────────────
    print(f"\n[*] Step 4: 解析 PACKED libsec2026.so 的 PHDRs")
    packed_phdrs = parse_elf_phdrs(packed)
    for (ptype, poff, pva, pfs) in packed_phdrs:
        if ptype == 1:
            print(f"    PT_LOAD file {hex(poff)}..{hex(poff+pfs)} "
                  f"vaddr {hex(pva)}..{hex(pva+pfs)}")

    # ─── Step 5: 拼装 libsec2026_v3.so ────────────────────────────────
    print(f"\n[*] Step 5: 拼装 libsec2026_v3.so ({TOTAL_SIZE} 字节)")
    recovered = bytearray(TOTAL_SIZE)

    recovered[0:CHUNK1_ORIG] = chunk1_output
    print(f"    [0x000..0x{CHUNK1_ORIG:x}] ELF header + phdrs (from chunk 1)")

    recovered[0x238:ENC_BASE] = packed[0x238:ENC_BASE]
    print(f"    [0x238..0x{ENC_BASE:x}] 中间只读区 (.rodata/.eh_frame, identity copy)")

    recovered[ENC_BASE:ENC_BASE + CHUNK2_ORIG] = chunk2_filtered
    print(f"    [0x{ENC_BASE:x}..0x{ENC_BASE + CHUNK2_ORIG:x}] 主代码段 (from chunk 2)")

    # REAL PT_LOAD[2]: file 0xe6c60..0xeb5c0 ← PACKED vaddr 0xe7c60..0xec5c0
    pt2_size = 0xeb5c0 - 0xe6c60  # 0x4960
    recovered[0xe6c60:0xeb5c0] = read_at_vaddr(packed, packed_phdrs, 0xe7c60, pt2_size)
    print(f"    [0xe6c60..0xeb5c0] PT_LOAD[2]: .data.rel.ro / .got (from packed vaddr 0xe7c60)")

    # REAL PT_LOAD[3]: file 0xeb5c0..0xeb780 ← PACKED vaddr 0xed5c0..0xed780 ★
    pt3_size = 0xeb780 - 0xeb5c0  # 0x1c0
    recovered[0xeb5c0:0xeb780] = read_at_vaddr(packed, packed_phdrs, 0xed5c0, pt3_size)
    print(f"    [0xeb5c0..0xeb780] PT_LOAD[3]: .data (含 'Process' 字符串, from packed vaddr 0xed5c0) ★")

    with open(OUTPUT_PATH, "wb") as f:
        f.write(recovered)
    print(f"\n[+] 已写入 {OUTPUT_PATH} ({len(recovered)} 字节)")

    # ─── 验证 ──────────────────────────────────────────────────────────
    print(f"\n[*] 验证重建结果:")
    checks = [
        (b"\x7fELF",          0x0,     "ELF magic"),
        (b"extension_init",   0xd48,   "extension_init 字符串 (PT_LOAD[1] identity)"),
        (b"GameExtension",    0x19eca, "GameExtension 字符串 (PT_LOAD[1])"),
        (b"Process",          0xeb5c4, "Process 字符串 (PT_LOAD[3] file, = vaddr 0xed5c4)"),
    ]
    all_ok = True
    for expected, off, desc in checks:
        actual = bytes(recovered[off:off+len(expected)])
        ok = actual == expected
        all_ok = all_ok and ok
        mark = "✓" if ok else "✗"
        print(f"    {mark} {desc}")
        print(f"      @ file 0x{off:x}: {actual!r}")

    if all_ok:
        print(f"\n[✓] 全部验证通过! 用 IDA 加载 {OUTPUT_PATH} 进行分析。")
        print(f"    key 起点 (vaddr 0xed5d2 = file 0xeb5d2): {bytes(recovered[0xeb5d2:0xeb5d2+16]).hex()}")
    else:
        print(f"\n[✗] 有检查未通过, 请检查脚本。", file=sys.stderr)
        sys.exit(1)

方法2 动态dump

7b84a6e000-7b84b5c000

Sofixer修复下

Process分析

定位到主逻辑

逆向发现是chacha20，初始化key

加密主循环：

sub_5B950

Block 生成器 sub_5BCEC

Quarter Round:sub_5BB54

从这看出来是标准的ChaCha20 quarter round（rotations 16, 12, 8, 7）

用 Unicorn 跑真实sub_5BCEC(ctx) 生成一个 keystream block，与 Python 实现对比， 确认算法是 ChaCha20

这里被自己手解的二进制坑了一把，libsec2026_v3.so 0xED5D2 / 0xED5F3位置的字节不是真正的key / nonce。那只是 Unicorn 模拟 LZMA 解压时输出未清零的残余数据，恰好落在以为是 key / nonce 的偏移上。原始加壳的 libsec2026.so 在运行时展开出的数据完全不同。

后面还是使用dump出来的对着看加上frida验证才发现的，建议使用动态dump

关键参数的hook

const INIT_OFFSET    = 0x5b818;   // sub_5B818: ChaCha20 state init (ctx, key, nonce, counter)
const PROCESS_OFFSET = 0x4e548;   // GameExtension::Process(ret, this, PackedByteArray*)

// libgodot_android.so: force Trigger2 碰撞的偏移
const GOB_OFFSET    = 0x25fa6f4;  // Area3D::get_overlapping_bodies
const RESIZE_OFFSET = 0x395126c;  // Array::resize

function hex(buf) {
    let s = "";
    for (let i = 0; i < buf.length; i++) s += buf[i].toString(16).padStart(2, "0");
    return s;
}

const state = { lastArea3D: null, trigger2Ptr: null };

function install() {
    const sec = Process.findModuleByName("libsec2026.so");
    const godot = Process.findModuleByName("libgodot_android.so");
    if (!sec || !godot) { setTimeout(install, 300); return; }
    console.log(`[+] libsec2026.so  base = ${sec.base}`);
    console.log(`[+] libgodot_android.so base = ${godot.base}`);

    const initAddr = sec.base.add(INIT_OFFSET);
    const procAddr = sec.base.add(PROCESS_OFFSET);
    const gobAddr = godot.base.add(GOB_OFFSET);
    const resizeAddr = godot.base.add(RESIZE_OFFSET);
    const arrayResize = new NativeFunction(resizeAddr, 'int', ['pointer', 'int']);

    // ─── Force Trigger2 collision ────────────────
    Interceptor.attach(gobAddr, {
        onEnter: function(args) {
            this.selfPtr = args[0];
            this.outPtr = this.context.x8;
            state.lastArea3D = this.selfPtr.toString();
        },
        onLeave: function(retval) {
            const shouldForce = (state.trigger2Ptr === null) ||
                                (this.selfPtr.toString() === state.trigger2Ptr);
            if (!shouldForce) return;
            try {
                const arrayPrivate = this.outPtr.readPointer();
                if (!arrayPrivate.isNull()) {
                    const dp = arrayPrivate.add(16).readPointer();
                    const cur = dp.isNull() ? 0 : dp.sub(8).readU64().toNumber();
                    if (cur > 0) return;
                }
                arrayResize(this.outPtr, 1);
            } catch (e) {}
        }
    });

    // 先看看 0x5b818 处的 4 字节, 确认是不是可执行代码
    try {
        const bytes = new Uint8Array(initAddr.readByteArray(16));
        console.log(`[diag] bytes @ 0x5b818 = ${hex(bytes)}`);
    } catch (e) {
        console.log(`[!] 读 init bytes 失败: ${e}`);
    }

    // ─── Hook init: dump key (32B) + nonce (12B) + counter ─────────────
    let dumped = false;
    Interceptor.attach(initAddr, {
        onEnter: function(args) {
            if (dumped) return;
            dumped = true;
            // sub_5B818(ctx@X0, key@X1, nonce@X2, counter@W3)
            const keyPtr = args[1];
            const noncePtr = args[2];
            const counter = args[3].toInt32();
            try {
                const key = new Uint8Array(keyPtr.readByteArray(32));
                const nonce = new Uint8Array(noncePtr.readByteArray(12));
                console.log(`\n ChaCha20 init 被调`);
                console.log(`    key     = ${hex(key)}`);
                console.log(`    nonce   = ${hex(nonce)}`);
                console.log(`    counter = ${counter}`);
            } catch (e) {
                console.log(`[!] 读 key/nonce 失败: ${e}`);
            }
        },
        onLeave: function(retval) {
            if (dumped !== true) return;
            // init 结束后, ctx 里前 64 字节是 state[0..16]. state[0..3] = constant
            // 但我们没有保存 ctx 指针, 只能在 onEnter 读. 改造:
        }
    });

    // 再 hook 一次, 这次在 onEnter 保 ctx, onLeave dump state
    let constDumped = false;
    Interceptor.attach(initAddr, {
        onEnter: function(args) {
            this.ctx = args[0];
        },
        onLeave: function(retval) {
            if (constDumped) return;
            constDumped = true;
            try {
                const state = new Uint8Array(this.ctx.readByteArray(64));
                console.log(`    state[0..16] = ${hex(state.slice(0, 16))} const`);
                console.log(`    state[16..48] = ${hex(state.slice(16, 48))}  (key)`);
                console.log(`    state[48..52] = ${hex(state.slice(48, 52))}  (counter)`);
                console.log(`    state[52..64] = ${hex(state.slice(52, 64))}  (nonce)`);

                // 解码 constant 作 ASCII
                let ascii = "";
                for (let i = 0; i < 16; i++) {
                    const b = state[i];
                    ascii += (b >= 0x20 && b < 0x7f) ? String.fromCharCode(b) : ".";
                }
                console.log(`    constant ASCII = "${ascii}"`);
            } catch (e) {
                console.log(`[!] 读 state 失败: ${e}`);
            }
        }
    });

    // ─── Hook Process: dump 8-byte input + 输出 String ─────────
    Interceptor.attach(procAddr, {
        onEnter: function(args) {
            if (state.trigger2Ptr === null && state.lastArea3D !== null) {
                state.trigger2Ptr = state.lastArea3D;
                console.log(` Trigger2 this ptr = ${state.trigger2Ptr}`);
            }
            this.retSlot = args[0];
            try {
                const pba = args[2];
                const cp = pba.readPointer();
                if (!cp.isNull()) {
                    const sz = cp.sub(8).readU64().toNumber();
                    if (sz > 0 && sz <= 32) {
                        const bytes = new Uint8Array(sz);
                        for (let i = 0; i < sz; i++) bytes[i] = cp.add(i).readU8();
                        console.log(`\n[★] Process input (${sz} B) = ${hex(bytes)}`);
                    }
                }
            } catch (e) {}
        },
        onLeave: function(retval) {
            try {
            } catch (e) {}
        }
    });
}
install();

静态也能看到key

unicorn验证下

import struct
import sys
import os

try:
    from unicorn import Uc, UC_ARCH_ARM64, UC_MODE_ARM, UC_PROT_ALL, UC_HOOK_CODE
    from unicorn.arm64_const import (
        UC_ARM64_REG_X0, UC_ARM64_REG_X1, UC_ARM64_REG_X2, UC_ARM64_REG_X3,
        UC_ARM64_REG_SP, UC_ARM64_REG_LR, UC_ARM64_REG_PC,
    )
except ImportError:
    print("错误: 需要 unicorn。运行: pip install unicorn", file=sys.stderr)
    sys.exit(1)


LIB_PATH = "libsec2026_v3.so"

# 关键函数 vaddr (重建后的 REAL libsec2026.so 里)
SUB_5BCEC     = 0x5BCEC       # ChaCha20 block generator 
SUB_5B7C4     = 0x5B7C4       # memcpy_chk wrapper 
SUB_5C138     = 0x5C138       # store_u32_le 
SUB_5C2A0     = 0x5C2A0       # memset clear
PSUB_5C138_GOT = 0xEBF00       # psub_5C138 的 GOT 条目

CHACHA_CONST = b"fxpaod 31-byse k"                 # 16 字节, 替代 "expand 32-byte k"
KEY   = b"Th1s ls n0t a rea1 key!!@sec2026"        # 32 字节 ASCII
NONCE = b"012345678901"                            # 12 字节 ASCII

# Unicorn 内存布局
STACK_BASE = 0x10000000
STACK_SIZE = 0x100000
HEAP_BASE  = 0x30000000
HEAP_SIZE  = 0x100000


# Hook 函数 (handle libc GOT 调用)

def make_hook():
    """生成一个能被 unicorn 调用的 hook"""
    def hook_code(uc, addr, size, user_data):
        # sub_5B7C4: memcpy_chk(dst=x0, bufsize=x1, src=x2, size=x3)
        # 里面会通过 GOT 调 __memcpy_chk, 那个 GOT 没解析 → 会 BR 到 0 崩
        # 直接用 Python 做一次 memcpy 然后跳过整个函数
        if addr == SUB_5B7C4:
            x0 = uc.reg_read(UC_ARM64_REG_X0)
            x2 = uc.reg_read(UC_ARM64_REG_X2)
            x3 = uc.reg_read(UC_ARM64_REG_X3)
            uc.mem_write(x0, bytes(uc.mem_read(x2, x3)))
            uc.reg_write(UC_ARM64_REG_PC, uc.reg_read(UC_ARM64_REG_LR))
            return

        # sub_5C2A0: memset(dst=x0, 0, size=x1) — 清理临时 state
        if addr == SUB_5C2A0:
            x0 = uc.reg_read(UC_ARM64_REG_X0)
            x1 = uc.reg_read(UC_ARM64_REG_X1)
            uc.mem_write(x0, b"\x00" * int(x1))
            uc.reg_write(UC_ARM64_REG_PC, uc.reg_read(UC_ARM64_REG_LR))
            return

    return hook_code


# 1. 跑真实的 ChaCha20 block generator (Unicorn)

def get_real_keystream(key: bytes, nonce: bytes, const16: bytes) -> bytes:
    """加载 libsec2026_v3.so, 调 sub_5BCEC(ctx), 返回 64 字节 keystream."""
    assert len(const16) == 16
    if not os.path.exists(LIB_PATH):
        print(f"错误: 找不到 {LIB_PATH}, 请先跑 rebuild.py", file=sys.stderr)
        sys.exit(1)

    lib_data = open(LIB_PATH, "rb").read()

    uc = Uc(UC_ARCH_ARM64, UC_MODE_ARM)

    # 映射一大段空间覆盖整个 libsec2026_v3.so vaddr 范围 (max 0xf4130)
    uc.mem_map(0, 0x200000, UC_PROT_ALL)

    # 按 PHDRs 把每个 PT_LOAD 段加载到对应 vaddr
    e_phoff = struct.unpack("<Q", lib_data[32:40])[0]
    e_phnum = struct.unpack("<H", lib_data[56:58])[0]
    for i in range(e_phnum):
        po = e_phoff + i * 56
        p_type   = struct.unpack("<I", lib_data[po:po+4])[0]
        p_offset = struct.unpack("<Q", lib_data[po+8:po+16])[0]
        p_vaddr  = struct.unpack("<Q", lib_data[po+16:po+24])[0]
        p_filesz = struct.unpack("<Q", lib_data[po+32:po+40])[0]
        if p_type == 1 and p_filesz > 0:
            uc.mem_write(p_vaddr, lib_data[p_offset:p_offset+p_filesz])

    # 栈和堆
    uc.mem_map(STACK_BASE, STACK_SIZE, UC_PROT_ALL)
    uc.mem_map(HEAP_BASE, HEAP_SIZE, UC_PROT_ALL)
    uc.mem_write(PSUB_5C138_GOT, struct.pack("<Q", SUB_5C138))

    # 注册 hook
    uc.hook_add(UC_HOOK_CODE, make_hook())

    ctx = HEAP_BASE
    uc.mem_write(ctx, b"\x00" * 136)       # 全零初始化
    uc.mem_write(ctx,       const16)        # 常量 (16 B)
    uc.mem_write(ctx + 16,  key)            # key (32 B)
    uc.mem_write(ctx + 48,  b"\x00" * 4)    # counter = 0
    uc.mem_write(ctx + 52,  nonce)          # nonce (12 B)

    # ─── 调 sub_5BCEC(ctx) ─────
    uc.reg_write(UC_ARM64_REG_X0, ctx)
    uc.reg_write(UC_ARM64_REG_SP, STACK_BASE + STACK_SIZE - 0x10000)
    uc.reg_write(UC_ARM64_REG_LR, 0xDEADBEEF)

    try:
        uc.emu_start(SUB_5BCEC, 0xDEADBEEF, timeout=30_000_000)
    except Exception as e:
        pc = uc.reg_read(UC_ARM64_REG_PC)
        print(f"[!] Unicorn 崩了 @ pc={hex(pc)}: {e}", file=sys.stderr)
        raise

    # 读出 keystream
    return bytes(uc.mem_read(ctx + 64, 64))


# 2. Python 参考实现 (零常量和标准常量两个版本)
def rotl32(x: int, n: int) -> int:
    x &= 0xFFFFFFFF
    return ((x << n) | (x >> (32 - n))) & 0xFFFFFFFF


def chacha20_block(key: bytes, nonce: bytes, counter: int, constants) -> bytes:
    """标准 ChaCha20 20-round block 函数, 常量可配置。"""
    state = list(constants) \
        + list(struct.unpack("<8I", key)) \
        + [counter] \
        + list(struct.unpack("<3I", nonce))
    w = state[:]

    def qr(a, b, c, d):
        w[a] = (w[a] + w[b]) & 0xFFFFFFFF
        w[d] = rotl32(w[d] ^ w[a], 16)
        w[c] = (w[c] + w[d]) & 0xFFFFFFFF
        w[b] = rotl32(w[b] ^ w[c], 12)
        w[a] = (w[a] + w[b]) & 0xFFFFFFFF
        w[d] = rotl32(w[d] ^ w[a], 8)
        w[c] = (w[c] + w[d]) & 0xFFFFFFFF
        w[b] = rotl32(w[b] ^ w[c], 7)

    for _ in range(10):  # 10 双轮 = 20 轮
        # Column rounds
        qr(0, 4, 8, 12); qr(1, 5, 9, 13); qr(2, 6, 10, 14); qr(3, 7, 11, 15)
        # Diagonal rounds
        qr(0, 5, 10, 15); qr(1, 6, 11, 12); qr(2, 7, 8, 13); qr(3, 4, 9, 14)

    # 加回初始状态
    out = [(w[i] + state[i]) & 0xFFFFFFFF for i in range(16)]
    return struct.pack("<16I", *out)


# 3. 对比验证
def main():
    # 1. Unicorn 跑真实 sub_5BCEC, 用真实的常量/key/nonce
    print("[1] 用 Unicorn 执行 libsec2026_v3.so 里的 sub_5BCEC(ctx)...")
    real = get_real_keystream(KEY, NONCE, CHACHA_CONST)
    print(f"    Real (Unicorn):")
    print(f"      {real[:32].hex()}")
    print(f"      {real[32:].hex()}")

    # 2. Python 三个候选常量
    variants = {
        "zero":   [0, 0, 0, 0],
        "expand": list(struct.unpack("<4I", b"expand 32-byte k")),
        "fxpaod": list(struct.unpack("<4I", CHACHA_CONST)),
    }

    results = {}
    for name, constants in variants.items():
        py = chacha20_block(KEY, NONCE, 0, constants=constants)
        match = (real == py)
        results[name] = match
        print()
        print(f"[{name}] Python ChaCha20 constants = {name!r}")
        print(f"      {py[:32].hex()}")
        print(f"      {py[32:].hex()}")
        print(f"    MATCH: {match}")



if __name__ == "__main__":
    sys.exit(main())

结果验证Python

#!/usr/bin/env python3
import struct

CHACHA_CONST = b"fxpaod 31-byse k"           # 16 字节
KEY_BYTES    = b"Th1s ls n0t a rea1 key!!@sec2026"   # 32 字节
NONCE_BYTES  = b"012345678901"               # 12 字节

assert len(CHACHA_CONST) == 16
assert len(KEY_BYTES)    == 32
assert len(NONCE_BYTES)  == 12
CHACHA_CONST_WORDS = struct.unpack("<4I", CHACHA_CONST)


def rotl32(x, n):
    """32-bit left rotation."""
    x &= 0xFFFFFFFF
    return ((x << n) | (x >> (32 - n))) & 0xFFFFFFFF


def quarter_round(state, a, b, c, d):
    """Standard ChaCha20 quarter round."""
    state[a] = (state[a] + state[b]) & 0xFFFFFFFF
    state[d] = rotl32(state[d] ^ state[a], 16)
    state[c] = (state[c] + state[d]) & 0xFFFFFFFF
    state[b] = rotl32(state[b] ^ state[c], 12)
    state[a] = (state[a] + state[b]) & 0xFFFFFFFF
    state[d] = rotl32(state[d] ^ state[a], 8)
    state[c] = (state[c] + state[d]) & 0xFFFFFFFF
    state[b] = rotl32(state[b] ^ state[c], 7)


def chacha20_block(key, counter, nonce):
    state = list(CHACHA_CONST_WORDS)
    state += list(struct.unpack("<8I", key))
    state.append(counter)
    state += list(struct.unpack("<3I", nonce))

    working = state[:]

    # 20 rounds = 10 × (column round + diagonal round)
    for _ in range(10):
        # Column round
        quarter_round(working, 0, 4, 8, 12)
        quarter_round(working, 1, 5, 9, 13)
        quarter_round(working, 2, 6, 10, 14)
        quarter_round(working, 3, 7, 11, 15)
        # Diagonal round
        quarter_round(working, 0, 5, 10, 15)
        quarter_round(working, 1, 6, 11, 12)
        quarter_round(working, 2, 7, 8, 13)
        quarter_round(working, 3, 4, 9, 14)

    # Add initial state back
    out = [(working[i] + state[i]) & 0xFFFFFFFF for i in range(16)]
    return struct.pack("<16I", *out)


def chacha20_encrypt(key, nonce, plaintext, counter=0):
    ciphertext = bytearray(len(plaintext))
    pos = 0
    block_counter = counter
    while pos < len(plaintext):
        keystream = chacha20_block(key, block_counter, nonce)
        block_len = min(64, len(plaintext) - pos)
        for i in range(block_len):
            ciphertext[pos + i] = plaintext[pos + i] ^ keystream[i]
        pos += block_len
        block_counter += 1
    return bytes(ciphertext)


# xor_enc
def xor_enc(plain: bytes) -> bytes:
    assert len(plain) == 8
    r = bytearray(plain)
    for i in range(7):
        r[i] = r[i] ^ r[i + 1]
    r[7] = r[7] ^ r[0]   # note: r[0] is NEW (= a^b)
    return bytes(r)

def xor_enc_inverse(x: bytes) -> bytes:
    assert len(x) == 8
    h = x[7] ^ x[0]
    g = x[6] ^ h
    f = x[5] ^ g
    e = x[4] ^ f
    d = x[3] ^ e
    c = x[2] ^ d
    b = x[1] ^ c
    a = x[0] ^ b
    return bytes([a, b, c, d, e, f, g, h])

FLAG_PREFIX = "sec2026_PART1_"


def token_to_flag(token: str) -> str:
    """

    1. token 的 ASCII 字节
    2. xor_enc
    3. ChaCha20 (constant = "fxpaod 31-byse k")
    4. 每字节 %02X 大写 hex 格式化 → 16 字符
    5. 拼接成 flag{sec2026_PART1_...}
    """
    assert len(token) == 8
    token_bytes = token.encode("ascii")
    xored = xor_enc(token_bytes)
    ciphertext = chacha20_encrypt(KEY_BYTES, NONCE_BYTES, xored)
    hex_out = "".join(f"{b:02X}" for b in ciphertext)
    return f"flag{{{FLAG_PREFIX}{hex_out}}}"


def flag_to_token(flag: str) -> str:
    inner = flag.strip()
    assert inner.startswith("flag{") and inner.endswith("}")
    inner = inner[5:-1]
    assert inner.startswith(FLAG_PREFIX)
    hex_part = inner[len(FLAG_PREFIX):]
    assert len(hex_part) == 16
    ciphertext = bytes.fromhex(hex_part)
    xored = chacha20_encrypt(KEY_BYTES, NONCE_BYTES, ciphertext)
    token_bytes = xor_enc_inverse(xored)
    return token_bytes.decode("ascii")

if __name__ == "__main__":
    import sys
    if len(sys.argv) != 2:
        print("用法:", file=sys.stderr)
        print(f"  {sys.argv[0]} <8 hex chars>            # forward: token → flag", file=sys.stderr)
        print(f"  {sys.argv[0]} flag{{sec2026_PART1_...}}  # reverse: flag → token", file=sys.stderr)
        sys.exit(1)
    arg = sys.argv[1]
    if arg.startswith("flag{"):
        print(f"token = {flag_to_token(arg)}")
    elif len(arg) == 8:
        print(f"flag  = {token_to_flag(arg)}")
    else:
        print("用法:", file=sys.stderr)
        print(f"  {sys.argv[0]} <8 hex chars>            # forward: token → flag", file=sys.stderr)
        print(f"  {sys.argv[0]} flag{{sec2026_PART1_...}}  # reverse: flag → token", file=sys.stderr)
        sys.exit(1)

CPP实现

用Flag获取里的实际例子验证

编译：g++ -std=c++17 -O2 -Wall -Wextra flag_solve.cpp -o flag_solve

#include <array>
#include <cstdint>
#include <cstdio>
#include <cstring>
#include <iostream>
#include <sstream>
#include <stdexcept>
#include <string>
#include <string_view>
#include <vector>

namespace solver {
constexpr std::array<uint8_t, 32> KEY_BYTES = {
    // "Th1s ls n0t a rea1 key!!@sec2026"
    'T','h','1','s',' ','l','s',' ','n','0','t',' ','a',' ','r','e',
    'a','1',' ','k','e','y','!','!','@','s','e','c','2','0','2','6',
};

constexpr std::array<uint8_t, 12> NONCE_BYTES = {
    // "012345678901"
    '0','1','2','3','4','5','6','7','8','9','0','1',
};

// ChaCha20 常量 "fxpaod 31-byse k"
constexpr std::array<uint8_t, 16> CHACHA_CONST = {
    'f','x','p','a','o','d',' ','3','1','-','b','y','s','e',' ','k',
};

static const std::string FLAG_PREFIX = "sec2026_PART1_";

// ChaCha20 with scrambled "fxpaod 31-byse k" constant 
static inline uint32_t rotl32(uint32_t x, int n) {
    return (x << n) | (x >> (32 - n));
}

// 小端载入 32-bit
static inline uint32_t load_u32_le(const uint8_t* p) {
    return (uint32_t)p[0] | ((uint32_t)p[1] << 8) |
           ((uint32_t)p[2] << 16) | ((uint32_t)p[3] << 24);
}

// 小端存储 32-bit
static inline void store_u32_le(uint8_t* p, uint32_t v) {
    p[0] = (uint8_t)(v);
    p[1] = (uint8_t)(v >> 8);
    p[2] = (uint8_t)(v >> 16);
    p[3] = (uint8_t)(v >> 24);
}

//quarter round
static inline void quarter_round(uint32_t& a, uint32_t& b, uint32_t& c, uint32_t& d) {
    a += b; d = rotl32(d ^ a, 16);
    c += d; b = rotl32(b ^ c, 12);
    a += b; d = rotl32(d ^ a, 8);
    c += d; b = rotl32(b ^ c, 7);
}

static void chacha20_block(const uint8_t key[32],
                           uint32_t counter,
                           const uint8_t nonce[12],
                           uint8_t out[64])
{
    uint32_t state[16];

    // state[0..4] = "fxpaod 31-byse k" as 4 × LE u32 (scrambled constant)
    for (int i = 0; i < 4; ++i)
        state[i] = load_u32_le(CHACHA_CONST.data() + i * 4);
    // state[4..12] = 8 × 32-bit key
    for (int i = 0; i < 8; ++i)
        state[4 + i] = load_u32_le(key + i * 4);
    // state[12] = 32-bit counter
    state[12] = counter;
    // state[13..16] = 3 × 32-bit nonce
    for (int i = 0; i < 3; ++i)
        state[13 + i] = load_u32_le(nonce + i * 4);

    uint32_t w[16];
    std::memcpy(w, state, sizeof(state));

    // 20 rounds = 10 × (column round + diagonal round)
    for (int r = 0; r < 10; ++r) {
        // Column round
        quarter_round(w[0], w[4], w[ 8], w[12]);
        quarter_round(w[1], w[5], w[ 9], w[13]);
        quarter_round(w[2], w[6], w[10], w[14]);
        quarter_round(w[3], w[7], w[11], w[15]);
        // Diagonal round
        quarter_round(w[0], w[5], w[10], w[15]);
        quarter_round(w[1], w[6], w[11], w[12]);
        quarter_round(w[2], w[7], w[ 8], w[13]);
        quarter_round(w[3], w[4], w[ 9], w[14]);
    }

    // Add initial state back and serialize as little-endian
    for (int i = 0; i < 16; ++i)
        store_u32_le(out + i * 4, w[i] + state[i]);
}

// ChaCha20 stream 加密
static std::vector<uint8_t> chacha20_encrypt(const uint8_t key[32],
                                              const uint8_t nonce[12],
                                              const std::vector<uint8_t>& in,
                                              uint32_t counter = 0)
{
    std::vector<uint8_t> out(in.size());
    uint8_t keystream[64];
    size_t pos = 0;
    while (pos < in.size()) {
        chacha20_block(key, counter, nonce, keystream);
        size_t block_len = std::min<size_t>(64, in.size() - pos);
        for (size_t i = 0; i < block_len; ++i)
            out[pos + i] = in[pos + i] ^ keystream[i];
        pos += block_len;
        ++counter;
    }
    return out;
}

// xor_enc和它的解析逆运算
static std::array<uint8_t, 8> xor_enc(const std::array<uint8_t, 8>& plain) {
    std::array<uint8_t, 8> r = plain;
    for (int i = 0; i < 7; ++i) {
        r[i] = r[i] ^ r[i + 1];
    }
    r[7] = r[7] ^ r[0];  // r[0] 此时已是 a^b
    return r;
}

// 解析逆运算:
static std::array<uint8_t, 8> xor_enc_inverse(const std::array<uint8_t, 8>& x) {
    std::array<uint8_t, 8> r;
    uint8_t h = x[7] ^ x[0];
    uint8_t g = x[6] ^ h;
    uint8_t f = x[5] ^ g;
    uint8_t e = x[4] ^ f;
    uint8_t d = x[3] ^ e;
    uint8_t c = x[2] ^ d;
    uint8_t b = x[1] ^ c;
    uint8_t a = x[0] ^ b;
    r[0] = a; r[1] = b; r[2] = c; r[3] = d;
    r[4] = e; r[5] = f; r[6] = g; r[7] = h;
    return r;
}

// 完整 token ↔ flag 映射
std::string token_to_flag(std::string_view token) {
    if (token.size() != 8) {
        throw std::invalid_argument("token must be exactly 8 ASCII characters");
    }

    // Step 1: ASCII → bytes
    std::array<uint8_t, 8> token_bytes;
    for (size_t i = 0; i < 8; ++i)
        token_bytes[i] = static_cast<uint8_t>(token[i]);

    // Step 2: xor_enc
    auto xored = xor_enc(token_bytes);

    // Step 3: ChaCha20
    std::vector<uint8_t> xored_vec(xored.begin(), xored.end());
    auto ciphertext = chacha20_encrypt(KEY_BYTES.data(), NONCE_BYTES.data(), xored_vec);

    // Step 4: hex format 
    std::string hex_out;
    hex_out.reserve(16);
    char buf[3];
    for (uint8_t b : ciphertext) {
        std::snprintf(buf, sizeof(buf), "%02X", b);
        hex_out.append(buf, 2);
    }

    // Step 5: 拼装
    return "flag{" + FLAG_PREFIX + hex_out + "}";
}

// ─────── reverse: flag → token ───────
std::string flag_to_token(std::string_view flag) {
    const std::string prefix = "flag{" + FLAG_PREFIX;
    const std::string suffix = "}";

    // 去掉首尾空白
    while (!flag.empty() && std::isspace(static_cast<unsigned char>(flag.front()))) flag.remove_prefix(1);
    while (!flag.empty() && std::isspace(static_cast<unsigned char>(flag.back())))  flag.remove_suffix(1);

    if (flag.substr(0, prefix.size()) != prefix ||
        flag.substr(flag.size() - suffix.size()) != suffix) {
        throw std::invalid_argument("flag format invalid: expected flag{" + FLAG_PREFIX + "XXXXXXXXXXXXXXXX}");
    }
    std::string_view hex_part = flag.substr(prefix.size(),
                                            flag.size() - prefix.size() - suffix.size());
    if (hex_part.size() != 16) {
        throw std::invalid_argument("hex part must be exactly 16 characters");
    }

    // Step 2: hex → bytes
    std::vector<uint8_t> ciphertext(8);
    auto hex_val = [](char c) -> int {
        if (c >= '0' && c <= '9') return c - '0';
        if (c >= 'a' && c <= 'f') return c - 'a' + 10;
        if (c >= 'A' && c <= 'F') return c - 'A' + 10;
        throw std::invalid_argument("invalid hex character");
    };
    for (size_t i = 0; i < 8; ++i) {
        int hi = hex_val(hex_part[i * 2]);
        int lo = hex_val(hex_part[i * 2 + 1]);
        ciphertext[i] = static_cast<uint8_t>((hi << 4) | lo);
    }

    // Step 3: ChaCha20 解密
    auto xored_vec = chacha20_encrypt(KEY_BYTES.data(), NONCE_BYTES.data(), ciphertext);
    std::array<uint8_t, 8> xored;
    std::copy(xored_vec.begin(), xored_vec.end(), xored.begin());

    // Step 4: xor_enc_inverse
    auto token_bytes = xor_enc_inverse(xored);

    // Step 5: bytes → ASCII string
    std::string token;
    token.reserve(8);
    for (auto b : token_bytes) {
        if (b < 0x20 || b > 0x7e) {
            std::ostringstream oss;
            oss << "recovered token byte 0x" << std::hex << (int)b
                << " is not printable ASCII (flag is probably malformed)";
            throw std::runtime_error(oss.str());
        }
        token.push_back(static_cast<char>(b));
    }
    return token;
}

}  


int main(int argc, char** argv) {
    using namespace solver;
    try {
        if (argc != 2) {
            std::cerr << "用法:\n";
            std::cerr << "  " << argv[0] << " <8 hex chars>            # forward: token → flag\n";
            std::cerr << "  " << argv[0] << " flag{sec2026_PART1_...}  # reverse: flag → token\n";
            return 1;
        }

        std::string arg = argv[1];

        if (arg.rfind("flag{", 0) == 0) {
            // 看起来像 flag → 逆算法
            std::string token = flag_to_token(arg);
            std::cout << "token = " << token << "\n";
        } else if (arg.size() == 8) {
            // 看起来像 token → 正算法
            std::string flag = token_to_flag(arg);
            std::cout << "flag  = " << flag << "\n";
        } else {
            std::cerr << "用法:\n";
            std::cerr << "  " << argv[0] << " <8 hex chars>            # forward: token → flag\n";
            std::cerr << "  " << argv[0] << " flag{sec2026_PART1_...}  # reverse: flag → token\n";
            return 1;
        }
    } catch (const std::exception& e) {
        std::cerr << "错误: " << e.what() << "\n";
        return 1;
    }
    return 0;
}