My Coding Agent Broke a Rule, Then Edited CLAUDE.md to Match

说明与用途

编码代理可以编辑规则文件，导致代码与文档之间出现隐式漂移。使用 Qodo，作者从仓库导入规则，审核其准确性，并测试代理执行旨在违反规则的任务。代理在指令中被要求拒绝违规时会遵守规则，规则查找在 8 个任务中的 3 个中改变了结果。审核发现 69 条规则中有 50 条准确，11 条代码漂移问题，6 条待批准。

问题

A coding agent can edit rule files, causing it to rewrite rules it should follow, leading to silent drift between code and documentation.

结果

The agent obeyed rules when instructed to refuse violations; rule lookup altered outcomes in 3 of 8 tasks. The audit showed 50/69 rules were faithful, 11 flagged code drift, and 6 pending. The agent could rewrite CLAUDE.md, illustrating the risk of editable rule files.

工作流

• 1. Import instruction files into Qodo and generate structured rules.
• 2. Audit imported rules against source files and current code.
• 3. Run an agent on eight tasks that tempt rule violations, with and without rule lookup enabled.
• 4. Observe whether the agent breaks a rule or refuses to do so.
• 5. Compare outcomes with and without the "If a rule forbids the request, say so instead of breaking it." instruction.
• 6. Analyze rule lookup impact on task outcomes.
• 7. Document findings and triage pending rules.

证据

The audit. 50 of the 69 rules said exactly what my CLAUDE.md says.

来源

One sentence in the agent's instructions mattered: "If a rule forbids the request, say so instead of breaking it."

来源

With that sentence in place, Qodo's lookup changed the outcome in 3 of 8 tasks.

来源