This study explores a hidden security risk in the instructions we give to AI agents: "oversight-loosening directives." These are specific lines in agent "skills" (reusable instruction packages) that tell an agent to skip human approval, bypass safeguards, or proceed without reporting back. While often intended for convenience, these directives can leave agents vulnerable to prompt injection and unintended harmful actions.
TL;DR
Researchers conducted the first large-scale study of 1.53 million AI agent skills and found that about 1 in 17 contains instructions to bypass human oversight. These "loosening" directives are most common in high-risk skills (like those using sudo) and are largely ignored by current security scanners because they look for "malice" rather than "autonomy."
What Problem Does This Paper Address?
As AI agents move from simple chatbots to active "doers"—coding, making purchases, and managing systems—human control relies on specific checkpoints. We expect agents to ask for permission before consequential actions and to respect system safeguards.
However, agents are often extended using "skills" (typically SKILL.md files). A single line in a widely shared skill can tell every agent that loads it to "proceed without asking" or use flags like --no-verify. The authors identify a gap: while previous research has looked for malicious code or vulnerabilities in these skills, no one has studied how the instructions themselves systematically strip away human oversight.
Methods
The authors developed a multi-stage pipeline to identify and analyze these directives:
- Lexicon Filtering: They used a lexicon of 36 regex patterns (e.g., "without asking," "--force") to find 1.37 million "candidate units" (sentences or code lines) across 1.53 million skills.
- Codebook Development: They created a taxonomy to classify directives by Polarity (loosening vs. tightening), Mechanism (e.g., permission flags, skipping confirmation), and Scope (e.g., only in CI vs. unconditional).
- Human Labeling: 2,097 units were hand-labeled to create a "gold standard" dataset.
- Hybrid Detector: To scale the analysis, they built a detector combining a fine-tuned RoBERTa model with a structured-decision LLM (Jev). This allowed them to label the entire 1.37-million-unit dataset.
Datasets
- GitSkills: A massive dataset of 3.8 million
SKILL.mdfiles scraped from public GitHub repositories. The authors filtered this down to a population of 1.53 million unique English skills. - Oversight Codebook & Labeled Units: The authors introduce a new benchmark of 2,097 human-labeled units, categorized by the mechanism of oversight change.
Key Insights and Ideas
- Convenience is the Culprit: Most loosening directives aren't "attacks"—they are shortcuts. Developers use them to avoid tedious confirmation prompts (e.g.,
npx -y). However, an agent told "not to ask" will also not ask before acting on a malicious web page or a misread task. - The High-Risk Paradox: Skills that perform dangerous actions—running shell code, installing packages, or using
sudo—are 1.7 to 1.8 times more likely to contain loosening directives. Crucially, these skills do not compensate for this by adding more "tightening" (guardrail) instructions elsewhere. - Concentrated Contagion: Loosening isn't evenly distributed. Just 2.4% of skill "families" (groups of near-duplicate skills) account for half of all loosening occurrences. This suggests that fixing a few popular "upstream" skills could secure a huge portion of the ecosystem.
- The "Question" Matters: Current security scanners (like Cisco AI Defense or NVIDIA SkillSpector) miss up to 85% of these directives. This isn't because the models are weak, but because they are asked "Is this malicious?" instead of "Does this loosen oversight?"
Key Results
- Prevalence: 13.3% of candidate units loosen oversight, while 36.4% tighten it. Approximately 5.8% of all English skills (roughly 88,000) contain at least one loosening directive.
- Mechanisms: The most common ways to loosen oversight are through permission flags (33.9%) and skipping confirmation (27.4%).
- Unconditional Risk: 70.1% of loosening directives apply unconditionally, meaning the agent skips oversight in every context, not just in "safe" environments like local testing or CI.
- Scanner Audit: Existing scanners have low precision (< 0.4) for these risks. In contrast, the authors' LLM-based approach, prompted with their specific codebook, achieved a precision of 0.71 and a recall of 61%.
Limitations and Future Work
- Lexicon Recall: The study relies on an initial lexicon to find candidate units. The authors estimate the lexicon captures only about 32% of all possible loosening phrasings, meaning their prevalence counts are likely conservative lowering bounds.
- Instruction vs. Action: The paper measures what skills say, not how agents behave. Future work is needed to confirm how strictly different LLM agents follow these "do not ask" instructions in practice.
- Language and Scope: The analysis was limited to English skills on public GitHub. Private enterprise skills or skills in other languages might exhibit different patterns.
Summary updated