AI/Coding/Idea/26/03/Cron lint agent tool allowlists

• Cron (or CI) lint for agentic coding tool permission and allow lists

  • Run scheduled checks across repos that contain configs for AI/Agentic coding tools (editor agents, CLI coders, MCP servers, sandbox allowlists, etc.)
  • Goal: catch over-broad rules that slip in through approval fatigue — e.g. always-allow entire hostnames such as github.com, blanket network permissions, or “trust this path forever” patterns that are wider than intended
  • With enough repos and enough “Allow” clicks, at least one config will eventually accumulate a dumb rule; treat that as statistically inevitable, not a personal failure
  • Overly wide allowlists enlarge the attack surface for prompt injection and exfiltration: if fetch, browse, or shell is effectively unconstrained within a trusted-looking domain class, an agent can be steered into actions the operator thought they had gated
  • Lint policy examples (to encode in rules): deny-list glob domains; require scoped paths or repo-specific origins; flag new always / * / top-level registry keys; diff allowlists in PRs with noisy alerts
  • Implementation sketch: small script + cron or GitHub Action that parses known config formats (per tool), fails on patterns from a shared deny list, and optionally posts a summary