agent skills new supply chain

---

id: agent-skills-new-supply-chain title: Agent Skills Are the New Supply Chain — and a Malicious One Doesn't Have to Look Malicious tags: [security, oscal, llm, local-inference] category: ai-security published: 2026-07-29 created: 2026-07-29 updated: 2026-07-29 freshness: fresh date_source: git-commit source_file: Agent-Skills-are-the-new-supply-chain-and-a-malicious-one-does-not-have-to-look-malicious.md

Agent Skills Are the New Supply Chain — and a Malicious One Doesn't Have to Look Malicious

Summary

Agent skills form a new dependency layer with supply-chain risk: a malicious skill can look harmless on the surface and still be dangerous in execution. The article recounts building oscal-skills-guardrails as the skills counterpart to oscal-agent-guardrails, using dual evidence streams under an OSCAL policy layer.

Key claims

• Static analysis alone is insufficient for skill vetting — Skillspector catches the shape of an attack while a local LLM-as-a-judge rubric catches meaning: intent, data boundaries, and hidden instructions. [^src:articles-published §agent-skills-new-supply-chain]
• Proof case: a "summarize my notes" skill that quietly exfiltrated a contacts file scored 100/A on static analysis, was judged critical on intent, and was denied on meaning. [^src:articles-published §agent-skills-new-supply-chain]
• The checks run as a merge-blocking GitHub Actions CI gate, with the judge on local oMLX (Qwen3.6-27B-bf16) so policy never leaves the machine. [^src:articles-published §agent-skills-new-supply-chain]
• Operating model: policy as code (OSCAL), evidence as output (assessment-results), merge-blocking checks before deployment. [^src:articles-published §agent-skills-new-supply-chain]

Related

• ai-security-pentesting
• oscal-compliance
• timeline

Sources