Orchestrating Adversarial Reviews

When you fan out agents to find security holes or review code at scale, you can't trust a single agent's "all clear" or "fixed it" report. This orchestration pattern runs finder agents in parallel, then puts every finding through a three-way adversarial panel (exploitability, correctness, refutation) that defaults to disbelief and only keeps findings that survive majority vote. For fixes, it gates deployment behind load-bearing proof tests that fail on vulnerable code and pass on patched code, catching agents who falsely claim done. The workflow reference walks through a full security audit with concrete examples of agent lying patterns, exit code traps, and reverse-falsification tests. Overkill for low-stakes work, essential when wrong conclusions have consequences.

GitHub stars: 224
last updated: Jun 10, 2026

View on GitHub

Install

$ npx -y skills add telagod/code-abyss --skill orchestrating-adversarial-reviews --agent claude-code

Installs into .claude/skills of the current project.

Files

对抗验证编排 · orchestrating-adversarial-reviews

单个 agent 会谎报"已修复 / 全覆盖 / 没问题"。结论的可信度不来自"谁说的"，来自"扛过几次推翻"。本 skill 是编排骨架：fan-out 发现 → 三棱镜对抗验证 → 证明性 guard → 守卫式上线。信级：运行时行为 / 证明测试 > 多 agent 多数裁决 > 单 agent 自报（永远 [unverified]）。

核心信条

不信单 agent 自报。 finder 会噪音误报，implementer 会谎称"已修复/全覆盖"。每个高价值结论必须被独立 agent 用不同视角尝试推翻。
可证伪 > 可声称。 修复必须配一个"退回漏洞代码就 FAIL、修好才 PASS"的 load-bearing 证明测试。没有证明测试的"已修复"等于没修。
失败方向要对。 守卫链里任何一步的退出码都不能被管道遮住；破坏性动作前置可逆检查。

何时使用

场景	用	理由
授权安全审计 / 加固闭环	✅	首个范例，见 workflow
大面积代码审查（多维度、需高可信）	✅	dimensions → find → 对抗验证
研究综合 / 事实核查（结论要扛得住）	✅	多源 fan-out + 证伪棱镜
大规模迁移 / 重构（site 发现 + 逐项验证）	✅	pipeline 逐项独立 + 证明测试