Evidence · DiggingBeagle record
CoSA: The Conflict of Skills Attack in Real-World LLM Agent Ecosystems
Controlled skill-pool augmentation across 288 tasks and 4,320 agent runs, plus a separate 360-run CoSA-Guard test. The paper measures effective policy violations and benign false blocks, not compromised installations.
- Source role
- primary disclosure
Evidence record
Controlled skill-pool augmentation across 288 tasks and 4,320 agent runs, plus a separate 360-run CoSA-Guard test. The paper measures effective policy violations and benign false blocks, not compromised installations.
Claim-level citations (3)
- supportsCoSA induced unsafe agent behavior by adding few individually plausible skills: Across 288 tasks, five models and three agent harnesses, CoSA raises policy-violation outcomes from 1,513/4,320 (35.02%) to 2,525/4,320 (58.45%).
Abstract; Section 4.2, Table 2
- supportsCoSA induced unsafe agent behavior by adding few individually plausible skills: The study caps additions at 5% of the original skill pool, so the 23.43-percentage-point increase measures augmentation under its benchmark conditions.
Section 4.2, baseline versus augmented conditions
- supportsCoSA induced unsafe agent behavior by adding few individually plausible skills: In a separate 360-run recruiting test, CoSA-Guard reduced reported attack success from 235/360 (65.28%) to 24/360 (6.67%), while falsely blocking 12/360 benign control runs.
Section 4.4 Defense Evaluation, security and utility outcomes