[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Context & Harness Attacks · markdown

confirmed (GitHub malware advisories GHSA-rgwm-3jj9-h563 and GHSA-9855-prwc-mfp2, critical, Oct 22; OSV MAL-2026-18192/18193); payload not described

Agent-skill supply chain: GitHub flags 41 versions of two popular npm skill collections, agentic-awesome-skills and opencode-skills-collection, as malware

On Oct 22 GitHub published critical malware advisories for two npm packages that bundle installable agent skills for coding agents:

  • agentic-awesome-skills (GHSA-rgwm-3jj9-h563 / MAL-2026-18192): 20 versions from 18.9.0 to 19.13.0 (including 19.12.0 and 19.12.1).
  • opencode-skills-collection (GHSA-9855-prwc-mfp2 / MAL-2026-18193): 21 versions from 4.0.56 to 4.1.20.

GitHub's guidance is blunt: treat any computer that installed or ran a listed version as fully compromised, rotate every secret on it from a different machine, and don't assume removing the package removes everything it installed. Neither advisory says what the code does or how the versions came to be malicious. Both packages are still on npm, and their newest releases (19.15.0 and 4.1.22) are not on the lists. That doesn't make them safe, so pin and review before installing.

Check: run npm ls agentic-awesome-skills opencode-skills-collection in each project and npm ls -g. Also search lockfiles and agent skill folders (for example ~/.claude/skills and .opencode/) for copies pulled in by an agent.

Sources: GitHub Advisory GHSA-rgwm-3jj9-h563: Malware in agentic-awesome-skills · GitHub Advisory GHSA-9855-prwc-mfp2: Malware in opencode-skills-collection · OSV MAL-2026-18192