Astra refusals decoded: safety training fires at the local level, not against safety research
thebasepoint · x · 2026-11-06
Responding to reports that Google's Astra refuses "safety research" requests, thebasepoint clarifies the refused prompts are locally unsafe requests whose ultimate goal is improving safety — Astra's actual safety training is applying at the local level, rather than the model rejecting safety research per se.
More from Models
- Kaggle launches Chart Arena, a crowdsourced leaderboard for catching AI chart crimes — MeganRisdal · 2026-11-06
- Dev turns a 15-page paper into an interactive course with Kimi K3.1 Pro in 10 minutes — Arindam_1729 · 2026-11-06
- Xing4.0-29B-A4B Weights Out, llama.cpp Support PR Already Submitted — pmttyji · 2026-11-06
- Dev claims audit found identity injections in Claude skills and plugins — DionysianAgent · 2026-11-06
- Cekura Bench launches voice AI benchmark spanning 50+ STT, TTS and speech-to-speech models — Scobleizer · 2026-11-06
- Weight-space surgery: Opus-driven edit cuts Qwen3.8-27B refusals with no capability loss — FeydRowan · 2026-11-06