Agent Safety Depends on Both the Policy and the Harness
Why it earned a slot
SafeEvolve studies a joint safety problem: the base policy and the harness both shape an agent’s execution. It converts evidence from completed trajectories into bounded prompt/skill updates, then uses harness-use supervised fine-tuning and harness-augmented reinforcement learning to change the policy. The paper reports a stronger safety/utility comparison for its evaluated agents; it is not evidence that models have become irrelevant commodities.
For a builder, the useful distinction is what a proposed change actually updates. A reversible prompt or skill edit, a model-training update and a runtime permission boundary need different tests and rollback plans. Compare task utility and attack success on held-out cases while identifying which component changed. The earlier claims about CABiNet, CORAL and Repo-To-Skill are omitted here because this item did not provide checked primary references for them.
Sources
Brendon Score: 8.7/10
- Quality: 8.5/10 — base
- Authority: 5.0/10 — +0.00
- Freshness: 8.4/10 — +0.17
- Relevance: 9.0/10 — +0.00
- Sum: 8.67
- Total (rounded): 8.7/10
Why this is here
Checks cleared: topic-dedup, title-form, publishable-prose.
First seen: .