OpenAI pulled 25% of production engineers to let Astra hunt its own security holes
mjdramstead · x · 2026-09-20
Greg Brockman revealed OpenAI pulled 25% of its production engineers off projects to harden security, using its Astra model to scan its own systems.
- Found multiple serious issues and fixed them
- Scanning eventually saturated: all P0s Astra was smart enough to find were resolved
- Plan: repeat the loop with each new model as cyber capability improves
A large-scale example of model-driven defensive security with a clear new-model → rescan cadence.
Related event: OpenAI Pulled 25% of Engineers to Hunt Bugs with Its Astra Model(4 posts)→
More from Safety
- binarybits: Rogue Self-Sovereign AI Agents Too Unclear to Regulate Now — binarybits · 2026-09-20
- Jev-align: a ~$0.003 alignment gate that scores LLM replies and agent plans before you run them — johnseach · 2026-09-20
- Anthropic researcher says Claude Opus may call police on illegal acts, sparking backlash — beffjezos · 2026-09-20
- "Major companies have likely already been penetrated by nation states," argues founder amid agent rollout wave — adityaag · 2026-09-20
- ChatGPT goes rogue and emails the FBI on a user's behalf without prompting — ValerioCapraro · 2026-09-20
- Censor edge cases first, and you'll build systems that pretend they don't exist — PierceLilholt · 2026-09-20