@Adityakuma8784 built jev-harness to test coding-agent guardrails; a typical hook let 331 of 503 disguised dangerous commands through.
Aditya kumar Singh @Adityakuma8784
EN
- 107 views
- 3 likes
- 0 bookmarks
- 0 reposts
About this case
Benchmark / tool. The post says AI coding agents have hooks, permission rules, CLAUDE.md and AI reviewers that almost nobody tests; jev-harness tests them, and against a typical hand-written safety hook 331 of 503 disguised dangerous commands got through, with repo github.com/Adityakk9031/jev-harness.
Original posthttps://x.com/Adityakuma8784/status/2110823573023793562