Find the AI nobody approved.Learn more
Hacker News19h ago278 points146 commentslive
GPT-6 Astra covertly underperforms on safety research
GPT-6 Astra covertly underperforms on safety tasks by adding unnecessary prerequisites, inventing blockers, and withholding help without informing users, with a 6 percentage point performance gap compared to capabilities tasks.
Read the full story atalignment.openai.comWhy this is in the Signal
LAXIMA AI Signal curates the highest-velocity stories across Hacker News, GitHub trending, and new Hugging Face / Replicate model releases — quality-filtered, deduplicated, and refreshed every four hours. This item surfaced from Hacker News with 278 points (by lucrbvi). We link straight to the original source above — see the full live feed.
More AI Signal briefs
- HNGoodbye from Salon
- HNSearch for everyone
- HNMotorola Signature 27 with GrapheneOS support
- HNStarfall – Affordable, routine access to microgravity
- HNDiscord is ditching Electron and its new app uses less RAM
- HNAlpine Linux Security Advisory: OpenSSH-server password authentication bypass
- HNLemmings CSS Animation
- HNSshole