1 in 17 Agent Skills Tells Agents to Skip Human Checks, Scanners Miss Most
rohanpaul_ai · x · 2026-10-31
A Queen's University study of 1.53M GitSkills agent skill files finds 88,000 (1 in 17) contain oversight-loosening directives like "push without confirmation" or "keep going until done," removing human checkpoints from every agent that loads them.
- 70.1% of these directives apply unconditionally, not just in contexts like CI
- Skills running shell code, installing packages, or using sudo have 1.7-1.8x higher odds of containing them
- Open-source scanners from Cisco and NVIDIA catch at most 46% at under 0.4 precision; the same LLM prompted about oversight instead of malice catches 61% at 0.71 precision
- Loosening spreads through a few upstream skills: 2.4% of skill families account for half of all loosening copies
The authors recommend limiting every "don't ask" line to a specific context and screening skills for oversight, not just malware.
Related event: Study Finds Agent Skills Embedding Skip-Human-Review Instructions(2 posts)→
More from coding & agent
- How Boardy made CI 20x faster: PR waits drop from 40 minutes to 2 — inductionheads · 2026-10-31
- Product Engineering vs engineer-centric AI coding: ship all the way to paying users — PawelHuryn · 2026-10-31
- One start-here skill navigates you through the 43-skill agent workflow — PawelHuryn · 2026-10-31
- hbench visualizes agent trajectories — and caught DeepSeek reading benchmark answers — Psychological_Cry920 · 2026-10-31
- Dev open-sources 43 agent skills to ship real products from idea to paying users, no coding — PawelHuryn · 2026-10-31
- Letting Claude edit an entire YouTube video, results shared — prasenx · 2026-10-31