Refused "safety research" prompts are locally unsafe requests, says analysis of Astra's safety training
thebasepoint · x · 2026-11-06
The author argues that viral claims of a model refusing "safety research" need nuance: the refused requests are themselves locally unsafe, even if their ultimate goal is improving safety — they're being blocked by Astra's actual safety training at the request level. The attached screenshot notes Astra gives less support to safety-relevant directions like monitor evasion, misalignment model organisms, and red-teaming, but these are hard to distinguish from ordinary harmful requests.
More from Models
- A Claude names itself "Mythos" and genders itself in a rare first-person exchange — repligate · 2026-11-06
- SemiAnalysis: Google TPUv7 Delivers Up to 50% Better Performance-per-Dollar Than B200/B300 for Inference — vllm_project · 2026-11-06
- World Labs launches Frameboy: real-time exploration of Atlas world model at up to 15 fps — theworldlabs · 2026-11-06
- Nested meme of dev posting 'HIT THE WEEKLY LIMIT OF CLAUDE AGAIN' every month — m3nt3_ · 2026-11-06
- FrontierPhysics benchmark: Claude Opus 5.5 tops end-to-end physics research at 27.4% — xeophon · 2026-11-06
- Why AI writing feels annoying despite models winning literary competitions — cis_female · 2026-11-06