AI
SkillDRE: Dual-Stage Red-Team Evolution of Agent Skills via Pre-Execution and Runtime Feedback
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 25/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 10:09:59 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 05:18:25 UTC.
Briefings
SkillDRE is a fully automated framework for evolving malicious skill packages through a dual-stage feedback loop. It constructs and validates a task-conditioned malicious objective alongside a verifiable judge rule, evolving the skill implementation while preserving legitimate task capability. Evaluated on SkillsBench across four victim models, the framework achieves an average attack success rate of 45.28%, exceeding the strongest baseline by 40.3%. The system combines scanner-guided evolution with runtime-guided refinement informed by execution outcomes, forming a cross-stage closed loop where each runtime revision returns to the pre-execution stage for rescanning and further optimization before re-execution.
Key points
- SkillDRE constructs and validates a task-conditioned malicious objective and a verifiable judge rule to evolve malicious skills through a dual-stage feedback loop.
- The framework achieves an average attack success rate of 45.28% on SkillsBench, exceeding the strongest baseline by 40.3% while preserving benign-task performance.
Context and limitations
Results are limited to SkillsBench evaluation across four victim models; pre-execution scanning and runtime defense stages were evaluated in isolation, and final submitted skills may introduce new scanner findings.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 10:09:59 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigadbiwd2zeedrvfh32phtq4edws4d6obmyxljwjdspk54gbmzwbi