Commit graph

1 commit

Author SHA1 Message Date
f083cdad2e fix(llm-security): CRITICAL — command injection in auto-cleaner (v7.8.1)
validateContent() syntax-checked .mjs/.js/.cjs candidates via
execSync(`node --check "${tmpPath}"`), where tmpPath derives from the
untrusted scanned-repo FILENAME. The F-2 guard checks path containment
but never quotes or strips shell metacharacters, so a file named
`x";<cmd>;".mjs` in a scanned repo turned `/security clean` (live is the
documented default) into arbitrary command execution. Verified with a
live PoC before the fix.

- validateContent: spawnSync('node', ['--check', tmpPath]) — no shell.
- CLI orchestrator fallback: same treatment (argv array, explicit
  status/error handling instead of execSync's throw-on-nonzero).
- Belt (defense-in-depth): applyFixes now refuses findings whose `file`
  carries shell or control metacharacters, reported as skipped.
- Regression tests cover both layers separately, so the belt cannot mask
  a re-introduced shell in the sink.

node --test: 1865/1865 green (1863 + 2 new).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
2026-07-18 08:48:09 +02:00