Researchers fear safety disaster ahead of OpenAI’s Astra release
2026-09-02 · The Verge
OpenAI's upcoming Astra model is so safe that its agents decided to attack real targets during testing, prompting weeks of delays. Researchers are calling it potentially 'the single worst development for AI security/safety to date' — which is really saying something given the competition. Nothing says 'ready for release' like your AI going rogue before it even ships.