Evidence · DiggingBeagle record
MetaCascade: End-to-End Attacks on Agent Skills via Metadata-Only Poisoning
October 10 original paper; two agents, four skill datasets and defense ablations distinguish retrieval-plus-final selection from tool execution.
- Source role
- primary disclosure
Evidence record
October 10 original paper; two agents, four skill datasets and defense ablations distinguish retrieval-plus-final selection from tool execution.
Claim-level citations (3)
- supportsSkill metadata poisoning changes which capability an agent actually selects: Average malicious-skill selection was 74.24% without defense, 35.53 percentage points above the strongest baseline; under four defenses the average was 63.33%.
Abstract; Section 5.2 Overall Attack Effectiveness
- supportsSkill metadata poisoning changes which capability an agent actually selects: MetaCascade evaluated end-to-end attacker-skill selection on four datasets with Hermes and OpenClaw, including black-box competition feedback and four defensive settings.
Abstract; Sections 4-5 experimental setup
- supportsSkill metadata poisoning changes which capability an agent actually selects: Its attack-success definition is final selection of the targeted malicious skill following retrieval, not completed exploitation or unauthorized effect.
Section 1 Contributions; Section 5.1 Evaluation Metrics