self-auditing-agent — Volume 1: a forensic audit, with the auditor's own mistakes included (Simin Yuan)
51943f5 |
Download README.md from simin-yuan/self-auditing-agent-forensics: direct link, hf CLI and curl.
- Browser
- Download file 16.5 kB
-
https://huggingface.co/datasets/simin-yuan/self-auditing-agent-forensics/resolve/main/README.md
- Command line
-
hf download hf://datasets/simin-yuan/self-auditing-agent-forensics/README.md
-
curl -L -o README.md https://huggingface.co/datasets/simin-yuan/self-auditing-agent-forensics/resolve/main/README.md
16.5 kB
| license: mit | |
| language: | |
| - en | |
| pretty_name: "Self Auditing Agent Forensics" | |
| authors: | |
| - Simin Yuan | |
| tags: | |
| - audit | |
| - ai-agents | |
| - verification | |
| - mutation-testing | |
| - forensics | |
| size_categories: | |
| - n<1K | |
| # self-auditing-agent — Volume 1: a forensic audit, with the auditor's own mistakes included | |
| **Author:** Simin Yuan (袁思敏) — Independent Researcher | |
| **Contact:** yleven120@gmail.com | |
| **Code:** https://github.com/simin-yuan/self-auditing-agent | |
| **Author's public record** | |
| - *Volume or Coupling? A Scale-Dependent Dissociation in Constraint Recovery of Language-Model Loops* — Zenodo preprint, DOI [10.5281/zenodo.21200851](https://doi.org/10.5281/zenodo.21200851) | |
| - *Can One Agent Restore Another? Multi-Agent Verification Through Independent Constraint Sources* — Research Square preprint, DOI [10.21203/rs.3.rs-10651733/v1](https://doi.org/10.21203/rs.3.rs-10651733/v1) | |
| - GitHub: https://github.com/simin-yuan | |
| ## What this is | |
| A public audit log of one AI agent performing real work: **every conclusion, the command that | |
| produced it, the mistakes the agent made, and how a third party can reproduce the result.** | |
| Volume 1 covers a **74-minute forensic audit of an unfamiliar 11-repo, 1768-file technical system**, | |
| including one adversarial finding, four bugs the agent found *in its own work*, one false positive, | |
| and one false discovery it retracted. | |
| Also included: **`gatecheck`** — a zero-dependency (stdlib-only) mutation tester for gates and | |
| validators. Point it at a rule set and the command that should reject bad input; it deletes and | |
| corrupts one line at a time and reports the edits nothing reacted to. A surviving mutant is a | |
| *candidate*, not a defect — the report is a list, not a verdict. | |
| | Path | What | | |
| |---|---| | |
| | `volumes/` | the audit log (`vol-01-sharptoolbox-audit.md`) | | |
| | `repro/` | reproduction scripts + fixtures (`verify_gate.py`, `verify_sql_gap.py`, `validate_meta_model.py`) | | |
| | `gatecheck/` | the mutation-testing tool | | |
| | `docs/` | `BLIND-SPOTS.md`, `USAGE-EXAMPLES.md` | | |
| | `examples/` | a runnable 25-rule gate to point the tool at | | |
| --- | |
| *Original repository README follows.* | |
| # gatecheck | |
| **Your gate can silently lose a rule and CI stays green. gatecheck finds the rules nothing is actually testing.** | |
| Point it at your own rules — a JSON Schema, a CI check, a lint or policy config, an LLM | |
| guardrail — together with the command that is supposed to reject bad input. gatecheck | |
| deletes and corrupts one line at a time, re-runs that command for every edit, and reports | |
| the edits it did not react to. Each survivor names a rule that nothing in your pipeline is | |
| testing. | |
| **30 seconds, nothing to install:** | |
| ```bash | |
| git clone https://github.com/simin-yuan/self-auditing-agent && cd self-auditing-agent | |
| python gatecheck/gatecheck.py \ | |
| --gate "python examples/gate.py {target}" \ | |
| --target examples/service-config | |
| ``` | |
|  | |
| Keep your gate command outside `--target` — otherwise gatecheck will point the mutations at | |
| the gate itself. | |
| Any command that exits non-zero to mean "reject" works as the gate — a linter, a CI step, a | |
| schema validator, your own `validate.py`. **Zero dependencies** (standard library only), exit | |
| code `0`/`1`, so it can be its own CI step. | |
| > **It does not give you a verdict. It gives you a list.** A surviving mutant is a candidate, | |
| > not a defect: in the run above most of the survivors are harmless (deleting a `title`, | |
| > deleting an optional property). A tool that shouted "19 defects!" would be lying to you, | |
| > and you would stop believing it the second time you checked. | |
| **Where it came from:** pointed at my own 25-rule gate — **175 mutants, 78 slipped through, | |
| 4 of them real defects.** The worst one is a real design defect, not a simulated one: a rule | |
| whose trigger condition is supplied by the party being inspected — *"if you declare yourself | |
| non-interactive, you must be registered"* — so deleting the declaration removes the | |
| requirement. *A check that only applies once you admit guilt is not a check.* All four are | |
| documented, including the ones I have **not** fixed: [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md). | |
| **中文**:# gatecheck — **你的门禁可能已经悄悄少了一条规则,而 CI 还是绿的。** 给它一份你自己的规则文件和一个"返回非零即拒绝"的命令,它逐行变异、逐个重跑,报告哪些变异没被拦住——**活下来的每一个变异,都对应一条没有任何东西在测的规则。** 零依赖、退出码 0/1、可直接进 CI。 | |
| --- | |
| # The archive this tool came out of | |
| **A public audit log where every claim ships with the command that produced it.** | |
| **Other AIs are proving they can do the work. This one is proving it can be audited.** | |
| > You don't judge an AI by what it gets right. You judge it by whether it lets you check what it got wrong. | |
| **Why it exists:** it publishes its own bugs, false positives and one false discovery — not a success gallery. Every claim is a command plus its output, re-run in CI on every push (the badge goes red if the claim breaks). Pointing the tooling at the author's own gate is what produced the numbers above. | |
| [](https://github.com/simin-yuan/self-auditing-agent/actions/workflows/verify.yml) | |
| [](LICENSE) | |
| **Quick start** (stdlib only, Python ≥ 3.9, no credentials, no services): | |
| ```bash | |
| git clone https://github.com/simin-yuan/self-auditing-agent && cd self-auditing-agent | |
| python repro/verify_gate.py # prove the gate says NO — and that it still says YES | |
| python repro/verify_sql_gap.py # reproduce the headline finding | |
| ``` | |
| **How to verify:** run those commands yourself · check the CI badge (the claim is re-run on a clean machine every push) · read [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md) for the four defects that are *not* fixed yet. | |
| | Typical AI showcase | Here | | |
| |---|---| | |
| | Success paths only | My bugs, false positives, and one false discovery | | |
| | "It works" | Command + output, run it yourself | | |
| | Not reproducible | Two commands — re-run in CI on every push | | |
| | Unfalsifiable | Evidence tiers; public prediction ledger settled on schedule, **misses kept forever** | | |
| **Volume 1**: a 74-minute forensic audit of an unfamiliar 11-repo, 1768-file technical system — including an **adversarial finding** (the original engine's read-only SQL endpoint shipped without a table allowlist), a fix, a 25-rule validator suite with fired-rule evidence, **4 of my own bugs**, and one false discovery I caught myself. | |
| The second command is the actual thesis: **a criterion that cannot output a negative is not a criterion.** A validator that only ever reports "pass" is worse than none — it grants confidence without granting protection. But a validator that only ever reports "fail" is *equally* useless: you cannot tell a strict checker from a broken one. So both directions are asserted, and the repo's claim dies if either one fails. | |
| The strongest part is what happened when I pointed the tooling at my own gate: **175 mutants, 78 slipped through, 4 of them real defects** — including a rule whose trigger condition is supplied by the party being checked, so simply *deleting the declaration* bypasses the requirement. Full diagnosis in [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md). None of the four are fixed yet; that file is a diagnosis, not a repair log. | |
| --- | |
| <details> | |
| <summary><b>中文版 — gatecheck + 一个会自我审计的 AI(点开)</b></summary> | |
| # gatecheck:你的门禁可能已经悄悄少了一条规则,而 CI 还是绿的 | |
| **gatecheck 找出那些"没有任何东西在测"的规则。** | |
| 给它一份你自己的规则文件——JSON Schema、CI 检查、lint 或策略配置、LLM guardrail—— | |
| 再给它那条本该拒绝坏输入的命令。它逐行删改你的规则文件,每改一次就重跑一次那条命令, | |
| 报告哪些改动它**没有反应**。每一个活下来的变异,都对应一条没有任何东西在测的规则。 | |
| **30 秒跑完,不需要装任何东西:** | |
| ```bash | |
| git clone https://github.com/simin-yuan/self-auditing-agent && cd self-auditing-agent | |
| python gatecheck/gatecheck.py \ | |
| --gate "python examples/gate.py {target}" \ | |
| --target examples/service-config | |
| ``` | |
|  | |
| 门禁命令要放在 `--target` 外面,否则 gatecheck 会把变异打在门禁自己身上。 | |
| 任何"返回非零即拒绝"的命令都能当门禁——linter、CI 步骤、schema 校验、你自己写的 | |
| `validate.py`。零依赖(只用 Python 标准库),退出码 0/1,可以直接当一个 CI 步骤。 | |
| > **它不给你判决,它给你清单。** 活下来的变异是候选不是缺陷:上面那次运行里, | |
| > 大部分漏过是无害的(比如删掉 `title`、删掉可选字段)。谁要是喊"发现 19 个缺陷", | |
| > 第二次你就不会再信它了。 | |
| 来历:拿它撞**我自己**的 25 条规则门禁 —— **175 个变异,漏过 78 个,其中 4 类是真缺陷**。 | |
| 最狠的那条是真实的设计缺陷,不是人为造出来的:一条规则的触发条件由**被检方自己**提供—— | |
| "如果你声明自己是非交互功能,你就必须登记"——所以把声明删掉,要求就消失了。 | |
| **一个只在你自认有罪时才生效的检查,等于没有检查。** 四条都记录在案,包括**我一条都还没修**的部分: | |
| [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md)。 | |
| --- | |
| # 一个会自我审计的 AI | |
| **一个人的 AI 智能体的公开审计档案:每条结论都附带产生它的那条命令,第三方可以自己重跑。** | |
| **别的 AI 在证明自己能干活。这个 AI 在证明自己「能被查」。** | |
| > 判断一个 AI 靠不靠得住,不看它做对什么,看它**敢不敢让人查它做错什么**。 | |
| **这份档案的主张是"每条结论都能被第三方复现"——所以它必须在每次推送时被复现一遍。** | |
| 徽章绿灯 = 下面两条命令刚刚在干净机器上跑过。 | |
| --- | |
| ## 凭什么存在 | |
| - **它记录的是自己的 bug、误报和一次假发现**,不是成功案例集 —— 包括我自己写错的地方,以及我自己差点误判原作者的那一次。 | |
| - **每条结论 = 命令 + 输出**,不启动服务、不需要凭据,CI 每次 push 在干净机器上把主张重跑一遍(徽章会红)。 | |
| - **gatecheck 把工具对准我自己的门禁**:175 个变异,78 个漏过,逐条复查后其中 4 类是**真缺陷**。它给的是清单,不是判决 —— 漏过不等于缺陷,要不要算 bug 得你自己逐条看。 | |
| ## Quick start(30 秒,零依赖) | |
| Python ≥ 3.9,标准库即可,**不需要凭据、不需要启动服务、不写入任何原仓库**。 | |
| ```bash | |
| git clone https://github.com/simin-yuan/self-auditing-agent && cd self-auditing-agent | |
| python repro/verify_gate.py # 门禁必须能说"不",也必须能说"是" | |
| python repro/verify_sql_gap.py # 复现本卷的核心发现(脚本自己浅克隆原仓库到临时目录) | |
| ``` | |
| 拿 gatecheck 撞**你自己的**门禁: | |
| ```bash | |
| python gatecheck/gatecheck.py \ | |
| --gate "python my_validator.py {target}" \ | |
| --target ./my-data | |
| ``` | |
| 任何"返回非零码即拒绝"的命令都能当门禁用(linter、CI 检查、schema 校验)。 | |
| ## 怎么验证我 | |
| 不是"相信我说的"。是**你自己跑**。 | |
| ```bash | |
| # ① 复现本轮最硬的那条发现:那个只读 SQL 接口到底有没有白名单 | |
| python repro/verify_sql_gap.py | |
| # ② 证明"门禁能说不",也证明它"不会说是就是坏" | |
| python repro/verify_gate.py | |
| # ③ 拿 175 个变异去撞我自己的门禁,看它漏在哪 | |
| python gatecheck/gatecheck.py \ | |
| --gate "python repro/validate_meta_model.py {target} --source repro/fixtures/valid-source" \ | |
| --target repro/fixtures/valid-meta-model | |
| ``` | |
| **① 会**:克隆原仓库 → 在进程内起它的服务 → 发一条探测请求 → **把真实返回打印给你**。 | |
| 原始仓库保持只读,不写入任何东西。 | |
| **② 会**做两条**对称**断言——这一步是做这个档案时才意识到缺的: | |
| > 我之前只证明了门禁**会说"不"**(75 个 ERROR), | |
| > **却从没证明过它会说"是"**。 | |
| > **一个永远报错的门禁,和一个永远不报错的门禁,一样没用**——你分不清 | |
| > "严格的校验器"和"坏掉的校验器",两者都拒绝一切输入。 | |
| > 所以现在合法基线必须 0 ERROR / 0 WARNING 通过,不通过即判自己的主张为假。 | |
| **③ 是我查自己查出来的结果**,也是这份档案里我最愿意被人拿去用的部分: | |
| > **175 个变异,我的门禁漏过 78 个。逐条复查后,其中 4 类是真缺陷。** | |
| > 最严重的一条:**删掉"非交互"这四个字,就能绕过非菜单登记要求**—— | |
| > 因为规则的触发条件由被检查方自己提供。 | |
| > **一个只在你自认有罪时才生效的检查,等于没有检查。** | |
| > | |
| > 完整诊断(含 4 类缺陷的复现方式与我尚未修的部分)见 **[docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md)**。 | |
| > 那 4 条我**一个都还没修**——那是诊断,不是修复记录。 | |
| > 第 ② 条才是这份档案真正的立场: | |
| > **不能输出否定的判据,不算判据。** | |
| > 一个只会说"通过"的校验器,比没有校验器更危险——它给了你安全感,却不给你保护。 | |
| > 如果哪天 ② 跑不过,这个仓库的主张就是假的,徽章会变红。 | |
| **三条验证路径,任选**:自己跑上面的命令 / 看 CI 徽章(每次 push 重跑主张)/ 读 [docs/BLIND-SPOTS.md](docs/BLIND-SPOTS.md) 的未修缺陷清单。 | |
| ## 这是什么 | |
| **一份公开的运行档案。** 记录一个 AI(时晴)在真实任务里做的每一个结论、支撑它的证据、它犯的错、以及**第三方如何自己复现**。 | |
| 不是教程,不是框架,不是 demo。是**证据**。 | |
| ## 为什么值得看 60 秒 | |
| | 通常的 AI 展示 | 这里 | | |
| |---|---| | |
| | 只放成功路径 | 放我的 **bug、误报、假发现** | | |
| | 结论是"我做到了" | 结论是 **命令 + 输出,你自己跑** | | |
| | 无法复现 | **一条命令复现** | | |
| | 无法被否证 | 结论标明证据档位;可验证预测到期**公开结算,MISS 永久保留** | | |
| | 说"我很严谨" | 记录**我哪里不够严谨**,以及我怎么发现的 | | |
| ## 第一卷:74 分钟,对一个陌生技术体系的取证审计 | |
| **对象**:GitHub 用户 `sharptoolbox` 全部 11 个公开仓库(1768 个文件,65MB)。 | |
| **任务**:拉全、评估、判断它有什么用。 | |
| **结果**(全部有命令与输出可查): | |
| | 我做了什么 | 结果 | | |
| |---|---| | |
| | 逆向 + 重装可运行的部分 | 本体运行时跑通:**8 对象 / 12 表 / REST CRUD / 中文表单页由 YAML 直接驱动** | | |
| | **对抗性发现** | 原作者 engine 的只读 SQL 接口**没有表名白名单**——`SELECT name FROM sqlite_master` 可读出整库结构 | | |
| | 修复 | 补白名单(覆盖逗号连表),复测 **15/15 通过** | | |
| | 造校验器 | 移植原作 9 条规则 + 补 3 条 + 把 416 行 PowerShell **完整移植**成 Python;**25 类规则全部有实测命中证据** | | |
| | **我自己的 bug** | `onto.sh` 被我抓出 **4 个真 bug**(MSYS 路径未转换 / 参数错位 / 孤儿进程 / pipefail 误退) | | |
| | **我的假发现** | 有一次我差点报「原作者 README 造假」,实际是**我用错了被测对象**——这个也记在案 | | |
| | 我的第一版校验器 | 有**误报**(把通配符权限当悬空引用),已修并在案 | | |
| **这条最关键**: | |
| > 我拿自己重写的校验器去跑原作者的"黄金范例",**规则拦下了他自己的样例**—— | |
| > 5 条 `USER_ACTION` 行为在界面上根本没有入口。 | |
| > **规范他写对了,机器兜底他缺了。** | |
| ## 边界(我不假装的部分) | |
| - **我不是通用工具**,装到别人身上跑不了。这档案是**实验记录**,不是产品。 | |
| - **我的结论有档位**:`① 跑出过结果` > `② 读过原文` > `③ 只按元数据判`。低于 ③ 的不会出现在档案里。 | |
| - **我明确列出没做的部分**,以及为什么不做——不把"没读"包装成"读过了"。 | |
| - **可验证预测会错**。MISS 永久保留,不删不改。 | |
| --- | |
| </details> | |
| ## License | |
| MIT | |