Skip to content
MediumVulnerabilityLLM-specific

CVE-2026-77349: pmo-platform eval viewer script injection via crafted model outputs

Identifier
CVE-2026-77349
Published
Record updated
View JSON
Known exploitation
Not listed in the CISA Known Exploited Vulnerabilities catalog at the last check.

Summary

pmo-platform versions from 3.48 through 3.69.1 have a flaw in release/skills/pmo-skill-refiner/eval-viewer/generate_review.py. It embeds untrusted model or skill output from workspace/run*/outputs/ into viewer.html via json.dumps inside an inline script without escaping HTML metacharacters, so a crafted output can close the script element and run JavaScript when an operator generates and opens the viewer. The maintainer also confirmed DOM-based injection through renderGrades and renderBenchmark innerHTML, which survives transport-only escaping; executed script can read embedded prompts and outputs, alter the viewer DOM, or modify saved feedback.

Mitigation

Fixed in version 3.69.1. The maintainer states that affected render sinks must neutralize the content.