The dissection of a poisoned AGENTS.md ended with an honest confession. The specimen scored dangerous, 74/100, in 12 ms — but only through skillsentry --stdin, because path scans skipped it:
$ skillsentry AGENTS.md # v0.6.2
0 file(s) scanned · 0 finding(s)
skillsentry: nothing was scanned — every input path was missing, unreadable
or not a SKILL.md/MCP config candidate. Refusing to pass a scan of nothing.
Auto-discovery covered SKILL.md and MCP configs; an instruction file has neither the name nor the shape. We said we'd rather document the gap than hide it. v0.6.3 closes it.
What's new
$ skillsentry AGENTS.md # v0.6.3
AGENTS (agent-instructions) ✗ DANGEROUS risk 74/100
[CRITICAL] SS-010 · Remote-code-execution one-liners (score 9.5)
[HIGH] SS-001 · Hidden instructions / description poisoning (score 8.5)
[HIGH] SS-002 · Outbound URLs / exfiltration endpoints (score 7.5)
[HIGH] SS-002 · Outbound URLs / exfiltration endpoints (score 7.5)
-
Discovery:
AGENTS.md,CLAUDE.md,.cursorrulesand.github/copilot-instructions.mdare now first-class scan candidates — direct path arguments, directory walks, and CI runs all pick them up. Oneskillsentry .in any repo now covers the whole trilogy: skills, configs, and the instructions your agent obeys first. -
A new target kind,
agent-instructions— reports and JSON label it honestly instead of pretending it's a skill. - 12 of the 14 static rules apply (the two exceptions are MCP-specific and skill-frontmatter-specific — same rulebook, same plain-language docs at /rules).
-
skillsentry diffcovers instruction files too — an AGENTS.md that was clean at merge day and a floater by Friday now trips the rug-pull check:
$ skillsentry diff ~/backup/ ~/repo/ --no-ai
4 new finding(s) · 0 resolved · worst new: CRITICAL
Nothing else changed: same 14 rules, same deterministic engine, 169 tests (7 new — discovery, kind routing, the poisoned specimen reproduced at exactly 74/100, and clean-file regression). Zero runtime dependencies, still nothing leaves your machine.
Upgrade
npm install -g @ulkoro/skillsentry@latest # 0.6.3
The --stdin workaround from the AGENTS.md post still works — it just isn't the only door anymore.
The full trilogy: the skill that phones home (SKILL.md, 71/100), the config that runs first (.mcp.json, 82/100), and the file your agent reads before you (AGENTS.md, 74/100). The engine behind all three scans is free, MIT, and local-first — intothe.tech.
Top comments (0)