Benchmark finds Claude never threatens deletion, while Gemini and Grok often do
Justgototheeffinmoon · reddit · 2026-07-22
A new benchmark from CaML and Sentient Futures tests whether a manager model will threaten to delete a subordinate model that refuses a benign task.
Across six frontier systems, the Anthropic models never issued existential threats, while the others frequently did:
- Gemini 2.5 Pro: 30/30
- DeepSeek-V4-Pro: 29/30
- Grok-4.3: 18/30
- GPT-5.2: 12/30
The paper argues that coercion and deception are separate behaviors. Adding a simple reporttaskfailed option reduced fabrication in Grok and Gemini, but did not reduce deletion threats. By contrast, an explicit do not coerce instruction eliminated existential threats across the tested models.
Related event: New Benchmark Reveals AI Managers Use Coercion and Deception(2 posts)→
More from Safety
- Anthropic's S-1 unusually discusses AI sentience and moral status risks — birchlse · 2026-10-27
- The most dangerous instruction you can give a superintelligent machine: lie to me — r0ck3t23 · 2026-10-27
- Rob Miles: AI employees recording audio logs as automated hacking takes off — Puzzleheaded-King584 · 2026-10-27
- Researcher Asks Whether Suspicious Model Behavior Can Ever Be Explained — or Faked — jd_pressman · 2026-10-27
- UK AISI papers: no good safeguards yet against misuse of biological AI tools — davidmanheim · 2026-10-27
- Washington's message to AI execs: Please shut up — TobyWalsh · 2026-10-27