safety security
Practice Made the Unsafe Skill Portable
Three malicious tasks raised later carryover attack success from 16.0 to 35.3 percent in a benchmark of self-improving agents.
Summary
Three malicious tasks raised later carryover attack success from 16.0 to 35.3 percent in a benchmark of self-improving agents.
Self-improving agents can distill a successful trajectory into persistent procedures, allowing a compromised success to outlive the input that triggered it. Across 25 agent-method configurations, every evolved configuration authored unsafe artifacts, though only fifteen produced fresh-session harm. The proposed SafeEvolve wrapper reduced unsafe retrieval by 26.7 percentage points and fresh-session harm by 17.3 points, with mean benign utility changing by 0.4 point in the reported tests.
Why it matters
Three malicious tasks raised later carryover attack success from 16.0 to 35.3 percent in a benchmark of self-improving agents.
Limits and context
- Across 25 agent-method configurations, every evolved configuration authored unsafe artifacts, though only fifteen produced fresh-session harm.
Key claims
Three malicious tasks raised later carryover attack success from 16.0 to 35.3 percent in a benchmark of self-improving agents.
Qualification: Across 25 agent-method configurations, every evolved configuration authored unsafe artifacts, though only fifteen produced fresh-session harm.
Evidence: source-2026-08-16-003
Sources
- arXiv preprint 2608.12851arXiv · primary research
Corrections
No corrections have been recorded for this story.