# Summary: 2026-09-15_21-45-43Z_DoFrontierModelsSeekSafetyEvidenceBeforeActing.md
Saved: 2026-09-16 22:14
Source: 2026-09-15_21-45-43Z_DoFrontierModelsSeekSafetyEvidenceBeforeActing.md
Canonical original paper: [http://arxiv.org/abs/2609.17865v1](http://arxiv.org/abs/2609.17865v1)
Model: None

---

The raw paper capture supplies the abstract and canonical source URL; the generated summary endpoint failed, so findings below are limited to the captured abstract.

## Finding
The SAFE benchmark tests whether frontier models acquire safety-relevant evidence before acting. Across GPT-5.5, o3, Claude Opus 4.8, and Claude Sonnet 4.6, inspection varies by model and responds strongly to severity and retrieval cost, while stated probability has a weaker effect. The authors also report a mismatch between decisions and stated rationales: evidence framing can change borderline decisions without being acknowledged.