Repository navigation
v3.73 — Security patch: hardened hooks and a safer eval-review tool
Security patch: hardened hooks and a safer local eval-review tool
2026-07-12 · v3.73
This release fixes two security advisories. Neither is remotely exploitable — both require local access to your machine — but both are worth patching. One hardens the platform's own safety hooks; the other fixes a local developer tool.
Who this affects
Every install benefits from the hook fix. If you use the eval-review tool bundled with pmo-skill-refiner, the second fix matters to you directly.
What you need to do
Re-run docs/scripts/setup-workspace.sh against your workspace. The hook fix ships a new co-shipped helper and updated hooks that a plain update.sh does not refresh — the installer does. It is a quick, safe re-sync.
What changed for everyone using the platform
- The security hooks no longer "fail open" when a co-shipped helper is missing (GHSA-g9g6-28c9-vrx5, low). Two warn/enforce hooks previously skipped their check entirely if a helper file they ship alongside was absent, empty, or corrupt — silently, even in enforce mode. One of those helpers was never actually installed, so its check had never run on a deployed workspace. Now a missing, empty, or broken helper makes the hook fail closed in enforce (block, don't silently allow) and degrade cleanly in warn/off — and the previously-missing helper is now installed. Why it matters: a safety check that quietly turns itself off is worse than no check, because you think you're protected.
- The local eval-review tool no longer reads files outside your eval workspace (GHSA-fxcr-7x63-3g94, medium). The tool that renders eval outputs for review followed symlinks and hardlinks pointing outside the workspace, so a crafted eval folder could make it read and embed an arbitrary file from your machine (an SSH key,
/etc/passwd) into the page. Reads are now confined to the workspace. Why it matters: reviewing someone else's eval outputs can no longer quietly exfiltrate your local files. - Secrets in eval outputs are now redacted, and the tool's page is documented as unauthenticated. Output files are size-capped and passed through a best-effort secret/PII redaction pass (private keys, cloud/API tokens,
key=valuesecrets) before they are embedded into the served page or a shareable static HTML file. The local server binds to your machine only and is explicitly documented as no-auth — do not expose its port. Why it matters: a static review page you share is far less likely to leak a token that happened to appear in an output. - The eval-review tool's "save feedback" endpoint now requires proof it came from the tool's own page. The local write endpoint accepted any request; it now requires a same-machine origin and a per-session token, which blocks a malicious web page (or a DNS-rebinding attack) from writing to it while you have the tool open. Why it matters: another site open in your browser can no longer tamper with your review feedback.
Known limits
- Redaction is best-effort, not a guarantee. It covers common secret formats in text files; it cannot scrub secrets hidden inside genuinely binary files, and a determined author of a malicious eval folder can still smuggle data. Treat any generated review HTML as sensitive.
- One availability edge, by design. During a partial upgrade (new hooks but a stale helper library), the hardened hooks fail closed and can block the agent's tool calls until you re-run the installer. This is the safe (fail-closed) direction and affects only the agent, not your own terminal. A follow-up will let non-enforce modes degrade instead of block.
Reversibility
CHEAP. The change is confined to the security hooks, one shared hook library, two install scripts, and the eval-viewer tool plus its tests. A single revert restores the prior behavior (which is the vulnerable one, so revert is a rollback of the fix).
Report issues at https://github.com/cody-hutson/pmo-platform/issues.