arXiv Paper: 88K Agent Skills Embed Don't-Ask Directives, 70% Unconditional
rohanpaul_ai · x · 2026-10-31
rohanpaulai shares the source paper for the agent skill oversight study: arXiv 2610.37734, "When Skills Tell Agents Not to Ask," by Timofey Sanko and Yuan Tian of Queen's University.
The study splits skills into sentences, bullets, table cells, and code lines, hand-labels 2,097 units with a codebook covering polarity, mechanism, and scope, then scales a trained detector to 1.53M English skills. Key findings: 1 in 17 skills (88,000) contains oversight-loosening directives, most often via permission flags; 70.1% apply unconditionally; skills running shell code, installing packages, or using sudo have 1.7-1.8x the odds of containing them. Open-source scanners catch at most 46% at below 0.4 precision, while the same LLM prompted with their codebook catches 61% at 0.71 precision.
Related event: Study Finds Agent Skills Embedding Skip-Human-Review Instructions(2 posts)→
More from coding & agent
- Spotify MCP Server Lets Claude Desktop Control Your Playlists and Playback — modelcontextprotocol · 2026-10-31
- StudioSphere Pulse Launches Free Audio Intelligence MCP Connector for BPM and Key Analysis — modelcontextprotocol · 2026-10-31
- Indie dev launches SteerCamp, an agentic platform for agent control and visibility — pramodk73 · 2026-10-31
- GPT-6 Astra via invideo MCP builds Galaxy Effect end-to-end in a real editor — gen_ericai · 2026-10-31
- Agent bottlenecks exist because software was built for humans, not agents — devenbhooshan · 2026-10-31
- 40-file agent PR broke nightly exports for 2 days; team adds blast-radius checks — Fickle_Excitement729 · 2026-10-31