
中文版
BLA(Blind LLM Agents)(GitHub:https://github.com/qybaihe/bla ,2026-10-25 创建,今晨抓取时已 5,008★)是一款开源的反 AI 逆向防御工具:专门保护你自己的软件发布产物,不被 AI 逆向 Agent(比如之前推过的 morluto/rea 这类自主反编译器和 LLM 侦察管线)拆解。
README 的口号是“Poison → Exhaust → Hallucinate. Make Reverse Engineering Bankrupt Again.”(下毒、耗尽、致幻,让逆向再次破产)。做法分三步:先“下毒”,往产物里埋格式合法的诱饵、假的 SGX/TPM 安全芯片委托清单、蜜罐 SourceMap 和影子状态机,让 Agent 在互相矛盾的假设里分神追假线索;再“耗尽”,精确触发分析器的资源上限(比如 ASAR 16 MiB 头、WASM 表示上限、SQLite 虚拟列求值陷阱),把 REA 的检查结果从 COMPLETE 打到 LIMITED 或 CapacityError;最后“纠缠”,用运行时 SHA-256 摘要把 JS、WASM、原生组件互相绑定,防止 Agent 剥掉诱饵或把导出函数当离线预言机调用。流程是 audit → protect → verify 三步走,14 个内置基准目标全部通过输入输出等价测试,模型估算的 token 放大倍数在 300×–2,800×。支持二进制、Electron 应用、JS 包、Python、WebAssembly 等多种产物形态,以 MCP 与 CLI 两种形态提供。作者在 README 里明确标注:仅限防御用途,保护你自己的知识产权。
需要注意:这是一个防御工具,不是帮你去逆向别人的东西——README 专门写了免责声明,强调只用于作者自己的发布产物。
为什么值得了解
- 它是 REA 浪潮的另一面:上个月 AI 逆向工具刚火起来,这个月防御工具就跟上了,攻防两边都在同一条赛道上狂奔
- 思路很有意思:不对抗模型的能力,而是攻击 Agent 的预算——让它在假线索上烧掉 token,在上限前撞墙
- 对独立开发者和小团队很实在:发一个 Electron 应用或 JS 库,最怕的就是被人用 AI 一键扒光,BLA 提供了一套现成的“发布前加固”流程
English Version
BLA (Blind LLM Agents) (GitHub: https://github.com/qybaihe/bla , created 2026-10-25, 5,008 stars at this morning’s fetch) is an open-source counter-AI-reverse-engineering defense tool: it protects your own software releases from being taken apart by AI reverse-engineering agents (such as morluto/rea, the autonomous decompiler featured earlier, and LLM recon pipelines).
The README’s motto is “Poison → Exhaust → Hallucinate. Make Reverse Engineering Bankrupt Again.” It works in three moves: first “poison”, planting format-valid canaries, fake SGX/TPM secure-enclave delegation manifests, honey SourceMaps and shadow state machines so agents split their attention across contradictory hypotheses and chase synthetic rabbit holes; then “exhaust”, tripping analyzers’ exact resource ceilings (an ASAR 16 MiB header, WASM representation caps, SQLite virtual-column evaluation traps) to downgrade REA’s verdict from COMPLETE to LIMITED or CapacityError; finally “entangle”, binding JS, WASM and native sidecars together with runtime SHA-256 digests so agents can’t strip the decoys or call exported functions as offline oracles. The workflow is audit → protect → verify, all 14 bundled benchmark targets pass input/output equivalence tests, and modeled token amplification runs 300×–2,800×. It covers binaries, Electron apps, JS bundles, Python and WebAssembly, and ships as both an MCP server and a CLI. The README states clearly: defensive use only, protecting your own intellectual property.
One caveat: this is a defense tool, not something to reverse-engineer other people’s software with — the README carries a disclaimer stressing it is only for the author’s own releases.
Why it’s worth knowing
- It is the other side of the REA wave: AI reverse-engineering tools went viral last month, and the defense tools caught up this month — offense and defense are sprinting on the same track
- The idea is clever: don’t fight the model’s capability, attack the agent’s budget — burn its tokens on fake leads and make it hit the ceiling before it reaches your core logic
- Very practical for indie developers and small teams: ship an Electron app or a JS library and the fear is someone stripping it bare with one AI command — BLA offers a ready-made “harden before release” flow

评论(0)
暂无评论