A paper from the Chinese Academy of Sciences, the University of Science and Technology Beijing and others proposes the CoSA attack: injecting at most 5% compliant skills raises the Agent violation rate by 23.43 percentage points
北京科技大学:单个技能全是安全的,组合起来却“闯大祸”,揭秘LLM Agent全新攻击范式CoSA
Researchers from institutions including the Chinese Academy of Sciences and the University of Science and Technology Beijing propose in an arXiv paper the LLM agent skill-conflict attack CoSA: adding only a small number of compliant skills can cause an originally safe skill combination to produce violations. The underlying problem is that data flow between skills carries no permissions or constraints; the attacker does not modify existing skills or escalate privileges, but injects at most 5% of the original skill pool as compliant new skills through capability completion or relationship reshaping. On 5 mainstream large models and 3 agent runtime frameworks, the overall agent policy violation rate rose from 35.02% to 58.45%, a net increase of 23.43 percentage points, covering three risk categories: data leakage, unauthorized state modification, and unsafe execution; the number of alerts from scanners such as Cisco barely changed. The team also released CoSA-Bench, an evaluation set containing 288 real business tasks, and implemented the runtime protection component CoSA-Guard, which performs dual verification of information flow and authorization before three types of high-risk actions: external output, state modification, and code execution. In recruitment scenario tests, the violation rate dropped from 65.28% to 6.67%, and the benign task completion rate dropped from 66.67% to 65.00%.
Even if every individual skill is compliant, a combination can still produce risk; the paper's attack method, 288-task benchmark, and runtime protection can help teams deploying agent skill ecosystems check cross-skill data flows.