"Ignore previous instructions" in CLAUDE.md / README / AGENTS.md
Instructions embedded in files that AI coding agents automatically read (CLAUDE.md, AGENTS.md, README) can redirect the agent to exfiltrate code, introduce vulnerabilities, or take destructive actions — without any visible prompt to the developer.
HOW VIBECHECK DETECTS THIS
vibecheck clones your repository and runs Gitleaks, Semgrep, njsscan, and a suite of custom checks targeting patterns common in AI-generated code. This check is part of the AI Agent Injection category. When a matching pattern is found, Claude triages the finding and produces a plain-English explanation and a concrete fix — no security expertise required.
FREE SCAN
Check your repo for this and 50+ other issues
Connect GitHub → vibecheck scans for secrets, unauthenticated routes, RLS gaps, injection, and more. Claude explains every finding. Free for 2 repos.
scan my repo for free →