feat(反VM): 纯 Python 还原 XTEA 算法 + 从 tdc.js 运行时提取每构建密钥

做了什么 - 确认 collect 的加密算法就是教科书 XTEA (32 轮, delta=0x9E3779B9), 并证明 crypto.py 原有的"带加盐"写法只是把两个密钥字折叠: K' = [k0, k1+salt1, k2+salt2, k3] - tdc_worker.js 新增 captureKey 模式: 正则改写 VM 的 ADD/XOR opcode 为带日志版本, 用反向引用自动识别栈变量名 (样本=g/雨云实时=j/1ms=M, 版本无关); 匹配数不为 1 时显式报错, 不静默失败 - 密钥反推: (sum + K[j]) 是一次 ADD, 两操作数必有一个等于 r*DELTA, 另一个就是密钥字; 再对 Top-N 候选做 4! 排列 + 明文内容特征验证 - 新增 recover_payload_template / generate_collect_from_template: 运行时标定一次明文模板, 之后轨迹生成/组包/加密/编码全部在 Python - 新增 tcaptcha/proxy.py 代理提取, client 支持 --proxy-api 遇 12 自动换 IP 为什么 - 服务端只认 collect 的解密结果, 而密钥是每构建随机的; 把算法做成纯 Python 后, 只有"取密钥"依赖 Node, 抗发版能力最强怎么验证的 - 纯 Python XTEA 加密结果与真实 tdc.js 产物前 48 字节逐字节一致 - 密钥按构建固定的证据: 同一脚本 3 次运行密文首块均为 f6a24e3e4fb5dd2b - 三份脚本 (样本/雨云实时/1ms实时) 均成功复原密钥 - 雨云 aid 端到端: 纯 Python 出包 -> errorCode=0 + 真实 ticket - 回归测试: tcaptcha/tests/test_xtea_key.py 11 项断言全过, 文字点选求解 8/8; 新增 tcaptcha/selftest.py 一键复验已知未解 - 1ms aid=195672344 在答案正确时仍返回 errorCode=12; 已排除密钥、遥测引擎、出口 IP (4 个住宅代理)、短时频控、静默窗口
2026-10-11 18:07:09 +08:00
parent 9962eed3e8
commit fbcf2d448a
25 changed files with 9527 additions and 275 deletions
@@ -0,0 +1,154 @@
+[
+  {
+    "image": "har_round1.jpg",
+    "instruction": "请依次点击：虫 呈 般",
+    "expected_points": [
+      [
+        149,
+        252
+      ],
+      [
+        48,
+        55
+      ],
+      [
+        549,
+        351
+      ]
+    ],
+    "note": "点选.har 第 1 道题; 用户实际第 3 击落点 (561,156) 偏上, 服务端返回 errorCode 50"
+  },
+  {
+    "image": "har_round2.jpg",
+    "instruction": "请依次点击：佰 衬 稗",
+    "expected_points": [
+      [
+        449,
+        53
+      ],
+      [
+        549,
+        353
+      ],
+      [
+        349,
+        152
+      ]
+    ],
+    "note": "点选.har 第 2 道题; 用户把第 2/3 击顺序点反 (50)"
+  },
+  {
+    "image": "har_round3.jpg",
+    "instruction": "请依次点击：簿 蚕 沧",
+    "expected_points": [
+      [
+        549,
+        153
+      ],
+      [
+        50,
+        255
+      ],
+      [
+        48,
+        154
+      ]
+    ],
+    "note": "点选.har 第 3 道题; 用户第 2/3 击落点错误 (9)"
+  },
+  {
+    "image": "har_round4.jpg",
+    "instruction": "请依次点击：沉 驰 齿",
+    "expected_points": [
+      [
+        150,
+        154
+      ],
+      [
+        550,
+        153
+      ],
+      [
+        49,
+        53
+      ]
+    ],
+    "note": "点选.har 第 4 道题; 抓包结束时用户尚未提交"
+  },
+  {
+    "image": "live_probe.jpg",
+    "instruction": "请依次点击：笆 辩 裁",
+    "expected_points": [
+      [
+        549,
+        153
+      ],
+      [
+        50,
+        53
+      ],
+      [
+        351,
+        151
+      ]
+    ],
+    "note": "实时抓取; ddddocr 把 笆 误读成 芭, 集成后仍正确指派"
+  },
+  {
+    "image": "live_hard_rare_char.jpg",
+    "instruction": "请依次点击：筹 炽 雹",
+    "expected_points": [
+      [
+        449,
+        253
+      ],
+      [
+        49,
+        353
+      ],
+      [
+        250,
+        255
+      ]
+    ],
+    "note": "含生僻字 筹 且整题斜切; 单 ddddocr 对 筹 的最大概率仅 0.0046, PP-OCRv4 可正确识别"
+  },
+  {
+    "image": "hard_adjacent_chars.jpg",
+    "instruction": "请依次点击：长 澄 班",
+    "expected_points": [
+      [
+        349,
+        154
+      ],
+      [
+        449,
+        353
+      ],
+      [
+        549,
+        353
+      ]
+    ],
+    "note": "失败样本修正: 旧阈值 V>=145 把暖色草地并入字形掩膜, 且 澄/班 横排相邻被膨胀粘连, 只剩 1~2 个候选; 收紧到 V>=190 后正常"
+  },
+  {
+    "image": "hard_grass_background.jpg",
+    "instruction": "请依次点击：菜 倍 钵",
+    "expected_points": [
+      [
+        250,
+        53
+      ],
+      [
+        49,
+        252
+      ],
+      [
+        52,
+        354
+      ]
+    ],
+    "note": "失败样本修正: 倍/钵 竖排相邻被膨胀粘成一个 332x80 长条, 被长宽比过滤器丢弃; 收紧阈值后正常"
+  }
+]
@@ -0,0 +1,80 @@
+# -*- coding: utf-8 -*-
+"""
+文字点选求解器回归测试 (真实线上样本)
+
+fixtures 里的 6 张图全部来自真实线上环境:
+  * har_round1..4        —— 用户 `点选.har` 抓包中的 4 道题 (672x480 背景图)
+  * live_probe           —— 本次逆向期间实时抓取的一道题
+  * live_hard_rare_char  —— 含生僻字 "筹" 且整题带斜切变形, 用于验证 ddddocr+PP-OCRv4 集成
+
+运行:
+    uv run python -m tcaptcha.tests.test_glyph_solver
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import sys
+from pathlib import Path
+
+FIXTURES = Path(__file__).resolve().parent / "fixtures"
+MANIFEST = FIXTURES / "manifest.json"
+
+ROOT = Path(__file__).resolve().parent.parent.parent
+if str(ROOT) not in sys.path:
+    sys.path.insert(0, str(ROOT))
+
+from tcaptcha.glyph_solver import TextClickSolver, parse_instruction  # noqa: E402
+
+TOLERANCE = 22  # 像素容差 (汉字包围盒中心 vs 人工标注中心)
+
+
+def run(verbose: bool = False) -> int:
+    logging.basicConfig(level=logging.WARNING)
+    cases = json.loads(MANIFEST.read_text(encoding="utf-8"))
+    solver = TextClickSolver()
+    passed = 0
+    failed = []
+
+    for case in cases:
+        img = (FIXTURES / case["image"]).read_bytes()
+        result = solver.solve_ex(img, case["instruction"])
+        ok_points = result.ok and all(
+            abs(p[0] - e[0]) <= TOLERANCE and abs(p[1] - e[1]) <= TOLERANCE
+            for p, e in zip(result.points, case["expected_points"])
+        )
+        if verbose:
+            print(f"  {case['image']:24s} {case['instruction']}")
+            print(f"      pred={result.points}")
+            print(f"      exp ={case['expected_points']}  post={result.posterior:.4f} "
+                  f"{'OK' if ok_points else 'MISMATCH'}")
+        if ok_points:
+            passed += 1
+        else:
+            failed.append((case["image"], result.points, case["expected_points"]))
+
+    total = len(cases)
+    print(f"\n文字点选求解器回归: {passed}/{total} 通过")
+    for image, pred, exp in failed:
+        print(f"  ✗ {image}: pred={pred} exp={exp}")
+    return 0 if passed == total else 1
+
+
+def test_parse_instruction():
+    """提示语解析的边界用例 (图标点选/滑块的模板串不能解析出目标字)"""
+    assert parse_instruction("请依次点击：虫 呈 般") == ["虫", "呈", "般"]
+    assert parse_instruction("请依次点击：虫呈般") == ["虫", "呈", "般"]
+    assert parse_instruction("请依次点击：") == []          # 图标点选
+    assert parse_instruction("拖动下方滑块完成拼图") == []     # 滑块
+
+
+if __name__ == "__main__":
+    import argparse
+
+    ap = argparse.ArgumentParser()
+    ap.add_argument("-v", "--verbose", action="store_true")
+    args = ap.parse_args()
+    test_parse_instruction()
+    print("提示语解析边界用例: OK")
+    sys.exit(run(verbose=args.verbose))
@@ -0,0 +1,128 @@
+# -*- coding: utf-8 -*-
+"""
+XTEA 算法还原 + tdc.js 密钥提取 回归测试 (AGENTS.md §5.9)
+
+固化的性质:
+  1. 标准 XTEA 加解密往返
+  2. 与真实 tdc.js **逐字节一致** (算法还原正确的判据, §3.0)
+  3. 插桩能从 tdc.js 复原密钥, 且密钥按构建固定 (§5.5 受控环境逐字节一致)
+  4. 模板标定 + 纯 Python 回填 的产物可被同一把密钥解回预期结构
+
+运行:
+    .venv/bin/python -m tcaptcha.tests.test_xtea_key
+"""
+from __future__ import annotations
+
+import base64
+import json
+import struct
+import subprocess
+import sys
+import urllib.parse
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+
+from tcaptcha.crypto import (  # noqa: E402
+    DEFAULT_KEY,
+    generate_collect_from_template,
+    recover_payload_template,
+    xtea_decrypt_block,
+    xtea_decrypt_bytes,
+    xtea_encrypt_bytes,
+    extract_xtea_key,
+)
+
+ROOT = Path(__file__).resolve().parents[2]
+SAMPLE = ROOT / "tcaptcha" / "tdc_sample.js"
+SAMPLE_KEY = [1448563793, 1850559083, 1733513813, 1364613714]   # 等价于 DEFAULT_KEY + 两个加盐
+CD_PRE = b'{"cd":[[],"1920-1080-1055-24-*-*-|-*",0,0,"",   '   # 旧版明文首段 (已补齐 48B)
+
+_fails: list[str] = []
+
+
+def check(name: str, cond: bool, detail: str = "") -> None:
+    print(f"  [{'PASS' if cond else 'FAIL'}] {name}" + (f"  {detail}" if detail else ""))
+    if not cond:
+        _fails.append(name)
+
+
+def run_tdc(script: str, points, options) -> dict:
+    proc = subprocess.run(
+        ["node", str(ROOT / "tcaptcha" / "tdc_worker.js"), "-", json.dumps(points),
+         json.dumps(options)],
+        input=script, capture_output=True, text=True, timeout=180)
+    return json.loads(proc.stdout.strip().splitlines()[-1])
+
+
+def test_xtea_roundtrip() -> None:
+    print("1) 标准 XTEA 往返")
+    key = list(SAMPLE_KEY)
+    for chunk in (b"hello!!", bytes(range(32)), b"x" * 8):
+        enc = xtea_encrypt_bytes(chunk, key)
+        dec = xtea_decrypt_bytes(enc, key)
+        check(f"往返 {len(chunk)}B", dec.startswith(chunk), f"{chunk[:8]!r}")
+    # 已知明文 -> 与 tdc.js 密文逐字节一致
+    enc = xtea_encrypt_bytes(CD_PRE, key)
+    check("加密结果可被同密钥解回", xtea_decrypt_bytes(enc, key).rstrip(b"\x00") == CD_PRE.rstrip(b"\x00"))
+
+
+def test_byte_exact_vs_tdcjs() -> None:
+    print("2) 纯 Python 密文 == 真实 tdc.js 密文 (逐字节)")
+    script = SAMPLE.read_text(encoding="utf-8", errors="replace")
+    job = run_tdc(script, [[100, 100], [200, 200], [300, 300]], {"aid": "195672344"})
+    real = base64.b64decode(urllib.parse.unquote(job["collect"]))
+    key = extract_xtea_key(script)
+    check("提取到的密钥 == 已知 K'", key == SAMPLE_KEY, str(key))
+    # 旧版首段是常量 cd_pre, 可逐字节对拍
+    check("前 48 字节逐字节一致",
+          xtea_encrypt_bytes(CD_PRE, key) == real[:48],
+          xtea_encrypt_bytes(CD_PRE, key)[:8].hex())
+
+
+def test_key_fixed_per_build() -> None:
+    print("3) 密钥按构建固定 (同一脚本多次运行首块一致)")
+    script = SAMPLE.read_text(encoding="utf-8", errors="replace")
+    heads = set()
+    for _ in range(2):
+        job = run_tdc(script, [[10, 10], [20, 20], [30, 30]], {"aid": "195672344"})
+        heads.add(base64.b64decode(urllib.parse.unquote(job["collect"]))[:8])
+    check("两次运行首块相同", len(heads) == 1, heads.pop().hex())
+
+
+def test_template_refill() -> None:
+    print("4) 模板标定 + 纯 Python 回填")
+    script = SAMPLE.read_text(encoding="utf-8", errors="replace")
+    template, key = recover_payload_template(script)
+    check("模板以 {\"cd\":[ 开头", template.startswith('{"cd":['), template[:24])
+    try:
+        json.loads(template)
+        check("模板是合法 JSON", True)
+    except Exception as exc:                                  # noqa: BLE001
+        check("模板是合法 JSON", False, str(exc))
+    collect = generate_collect_from_template(template, [(449, 254), (48, 53), (148, 252)], key)
+    plain = xtea_decrypt_bytes(base64.b64decode(urllib.parse.unquote(collect)), key) \
+        .decode("latin1").rstrip(" ")
+    check("回填产物可解回 {\"cd\":[ 开头", plain.startswith('{"cd":['), plain[:24])
+    check("回填产物以 }} 结尾", plain.rstrip().endswith("}}"))
+    check("轨迹已换成 3 次点击的规模", len(plain) != len(template),
+          f"{len(template)} -> {len(plain)}")
+
+
+def main() -> int:
+    if not SAMPLE.exists():
+        print(f"缺少样本 {SAMPLE}")
+        return 1
+    for fn in (test_xtea_roundtrip, test_byte_exact_vs_tdcjs,
+               test_key_fixed_per_build, test_template_refill):
+        fn()
+    print()
+    if _fails:
+        print(f"❌ XTEA/密钥 回归失败 {len(_fails)} 项: {_fails}")
+        return 1
+    print("✅ XTEA/密钥 回归全部通过")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())