Evidence · DiggingBeagle record
CoRedSkill: Stealthy Skill Composition Attacks via AST-Guided Injection
Original controlled study of multi-stage skill-package attacks using AST-guided modifications, measured on 150 custom benchmark cases.
- Published
- Oct 14, 2026
- Publisher
- Zhao, Zhai, Shen, Wu and Zhang / arXiv
- Source role
- primary disclosure
Evidence record
Original controlled study of multi-stage skill-package attacks using AST-guided modifications, measured on 150 custom benchmark cases.
Claim-level citations (3)
- supportsComposed agent skills can hide unauthorized operations while completing tasks: The authors evaluated 150 benchmark cases spanning 50 tasks and three unauthorized-operation objectives: exfiltration, tampering and privilege violations.
Abstract, benchmark construction and objective list
- supportsComposed agent skills can hide unauthorized operations while completing tasks: In the authors' offline benchmark, CoRedSkill reports 98.0% attack success, 95.3% legitimate task completion and 90.7% joint attack/task/audit-evasion success.
Abstract, main evaluation reported rates
- supportsComposed agent skills can hide unauthorized operations while completing tasks: CoRedSkill reports a 3.3% detection rate under the auditor used in its 150-case evaluation, alongside its separate attack-success and legitimate-task-success measurements.
Abstract, main evaluation, audit detection rate