SkillAtlas Publishes 3,014 Agent-Skill Attack Cases, and 42.5% Only Succeed After a Failed First Round
arXiv 2609.13353·medium signal
SkillAtlas converts private agent-skill security report bundles into reviewed, redacted, searchable public cases: 3,014 cases, 6,589 traces, 151,131 steps, 233 affected skills and 8 risk categories. The finding that matters for defenders is that 42.5% of successful cases first become successful after a non-success initial round, so a single sandbox run or a stable signature will miss them. Trajectory-grounded labels raise pre-execution guard accuracy to 0.770.