safety security
Safety Experience Rewrote Both Harness and Policy
SafeEvolve turns trajectory evidence into reversible harness updates, then trains the agent to use them.
Published Updated Story ID: mp-2026-09-03-016
Summary
SafeEvolve turns trajectory evidence into reversible harness updates, then trains the agent to use them.
On AgentDojo with Qwen3.5-4B, the authors report a threefold attack-success reduction while benign utility rose from 59.79% to 61.86%. This is a benchmark result for one co-evolution setup.
Why it matters
SafeEvolve turns trajectory evidence into reversible harness updates, then trains the agent to use them.
Limits and context
No additional limitation was separately recorded.
Key claims
SafeEvolve turns trajectory evidence into reversible harness updates, then trains the agent to use them.
Evidence: source-2026-09-03-018
Sources
- arXiv preprint 2609.02786arXiv · primary research
Corrections
No corrections have been recorded for this story.