Improve human output and contextualize Chromium sandbox setuid #5

New issue

Closed

opened 2026-08-29 03:38:13 +00:00 by hephaistos · 2 comments

hephaistos commented

2026-08-29 03:38:13 +00:00

Owner

Observed behavior

A routine brave-bin update is hard-blocked because chmod 4755 "$pkgdir/opt/brave-bin/chrome-sandbox" triggers the generic high-risk setuid rule. Luna correctly identifies the well-known Chromium sandbox helper, but policy intentionally prevents the LLM from lowering static risk. Human output is also difficult to scan and always emits a very long full-result JSON line.

Intended behavior

Keep unknown setuid/setgid and file-capability changes at high risk.
Classify the exact static pattern that installs $pkgdir/.../chrome-sandbox with mode 4755 as medium review, never automatic low/allow. This keeps a human gate without a false hard block.
Present package, model verdict, findings, decision, action, and report path in a structured human-readable layout.
Keep a stable compact machine-result line in human mode; full structured data remains available through --json and the saved report.
When medium review is requested, show the reasons before asking for confirmation.

Verification

Add regression tests for the Brave/Chromium case, generic setuid behavior, readable rendering, compact machine output, review prompt context, and existing exit-code/schema guarantees.

## Observed behavior A routine `brave-bin` update is hard-blocked because `chmod 4755 "$pkgdir/opt/brave-bin/chrome-sandbox"` triggers the generic high-risk setuid rule. Luna correctly identifies the well-known Chromium sandbox helper, but policy intentionally prevents the LLM from lowering static risk. Human output is also difficult to scan and always emits a very long full-result JSON line. ## Intended behavior - Keep unknown setuid/setgid and file-capability changes at high risk. - Classify the exact static pattern that installs `$pkgdir/.../chrome-sandbox` with mode `4755` as medium review, never automatic low/allow. This keeps a human gate without a false hard block. - Present package, model verdict, findings, decision, action, and report path in a structured human-readable layout. - Keep a stable compact machine-result line in human mode; full structured data remains available through `--json` and the saved report. - When medium review is requested, show the reasons before asking for confirmation. ## Verification Add regression tests for the Brave/Chromium case, generic setuid behavior, readable rendering, compact machine output, review prompt context, and existing exit-code/schema guarantees.

hephaistos commented

2026-08-29 03:38:13 +00:00

Author

Owner

Pris en charge manuellement par Hephaistos dans cette session.

Pris en charge manuellement par Hephaistos dans cette session.

hephaistos referenced this issue from a pull request that will close it,

2026-08-29 03:52:22 +00:00

Make audit output readable and review Chromium sandbox setuid #6

hephaistos commented

2026-08-29 03:52:52 +00:00

Author

Owner

Correctif proposé dans la PR #6 : #6

Le cas exact, à cible unique, brave-bin + chmod 4755 "$pkgdir/opt/brave-bin/chrome-sandbox" devient une revue medium avec confirmation explicite, pas un allow silencieux. Les autres opérations setuid/setgid/setcap restent high et bloquantes. La sortie interactive est structurée et compacte; une sortie redirigée conserve le résultat machine v1 complet, et les détails restent aussi dans le rapport et via --json.

Vérifications : 35 tests passent, syntaxe des PKGBUILD validée, signature Forgejo du commit vérifiée.

Correctif proposé dans la PR #6 : https://git.2027a.net/2027a/paru-llm-audit/pulls/6 Le cas exact, à cible unique, `brave-bin` + `chmod 4755 "$pkgdir/opt/brave-bin/chrome-sandbox"` devient une revue **medium** avec confirmation explicite, pas un allow silencieux. Les autres opérations setuid/setgid/setcap restent **high** et bloquantes. La sortie interactive est structurée et compacte; une sortie redirigée conserve le résultat machine v1 complet, et les détails restent aussi dans le rapport et via `--json`. Vérifications : 35 tests passent, syntaxe des PKGBUILD validée, signature Forgejo du commit vérifiée.