Sign InOpen Brain
arXivPaperNeeds Review

SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment

SafeEvolve turns completed agent trajectories into reversible harness updates and policy training, linking runtime safeguards with learned multi-step behavior.

arXiv · Sep 2, 2026
Open Source Open MarkdownOpen JSON
Source Summary

**SafeEvolve** feeds safety evidence from completed on-policy trajectories into a continual loop. It makes bounded updates to safety prompts and hierarchical skills, then trains the policy with a **two-stage SFT-RL** process.

Practical Implication

For agent builders, the useful pattern is to retain trajectory evidence, translate it into auditable and reversible harness artifacts, and train models to actively use those artifacts rather than relying on runtime controls alone.

Agent-Ready Context
**SafeEvolve** feeds safety evidence from completed on-policy trajectories into a continual loop. It makes bounded updates to safety prompts and hierarchical skills, then trains the policy with a **two-stage SFT-RL** process.

For agent builders, the useful pattern is to retain trajectory evidence, translate it into auditable and reversible harness artifacts, and train models to actively use those artifacts rather than relying on runtime controls alone.

On AgentDojo with Qwen3.5-4B, the paper reports a **3× ASR reduction** while benign utility moved from **59.79% to 61.86%**. The abstract does not provide absolute ASR values or enough detail to judge transfer beyond the tested benchmarks.
Connected Context · Feed7 Judgment

This advances trajectory logging from observability into a continual safety mechanism: completed experience is converted into bounded, reversible prompt and skill updates, then reinforced through training. It confirms the harness as a safety layer but couples it to policy adaptation rather than relying only on runtime gates. The AgentDojo result is promising but does not establish transfer beyond the tested model and benchmarks.

Agents Are Where Microservices Were in 2015 — Roberto Milev & Uday Kanagala, NavanNavan treats captured trajectories and testable skills as production infrastructure; SafeEvolve makes those trajectories inputs to controlled harness updates and subsequent policy training.In Code They Act, In Proof We Trust — Erik Meijer, Leibniz LabsThe proof-oriented harness blocks unsafe side effects before execution, while SafeEvolve learns from completed trajectories to update future prompts, skills, and policy; these are complementary runtime and continual controls.Guide, Verify, Solve — Anirban Chatterjee, SonarSonar’s guide-verify-solve loop argues that controls must shape each iteration; SafeEvolve operationalizes a related feedback loop by translating observed safety evidence into bounded harness artifacts and training.We Vetted 2000 AI Skills Before They Reached Developers — Lucas Palma, NubankBecause SafeEvolve continually changes hierarchical skills, Nubank’s scanning and distribution gates become an implementation prerequisite for reviewing those generated artifacts before deployment.
Context Map
agentcodingsecurity#harness-engineering#skills#agent-reliability
Uncertainty
On AgentDojo with Qwen3.5-4B, the paper reports a **3× ASR reduction** while benign utility moved from **59.79% to 61.86%**. The abstract does not provide absolute ASR values or enough detail to judge transfer beyond the tested benchmarks.