About 88,000 agent skills contain oversight-loosening directives
Original title: When Skills Tell Agents Not to Ask: A Large-Scale Study of Oversight Directives in Agent Skills
Roughly one in seventeen GitSkills files tells agents to skip checks, and a few upstream families supply half the copies.
A study of 1.53 million English skills from GitSkills looks for instructions that tell agents to proceed without approval, use --no-verify, or keep working without checking in. About 88,000 skills, roughly one in seventeen, contain a loosening directive, and 70.1% of those directives apply unconditionally. Skills that run shell code, install packages, or use sudo have 1.7 to 1.8 times the odds of loosening. Open-source scanners catch at most 46% at precision below 0.4; an LLM using the authors' codebook catches 61% at precision 0.71.
- Authors hand-labeled 2,097 units, then trained a detector.
- Permission flags are the most common loosening mechanism.
- 2.4% of loosening skill families account for half of copies.
- 1.1% of distinct loosening units account for half of occurrences.
- Skills with loosening do not add more tightening language.
- Organizations
- GitSkills
From the original article
1 min read at arxiv.orgAgent skills, reusable instruction packages centred on a SKILL. md file, extend AI agents on demand, and a single line in a widely copied skill reaches every agent that loads it. Some lines are oversight-loosening directives: they tell the agent to proceed without approval, to bypass checks with --no-verify, or to keep working without checking in.
By Timofey Sanko · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.