Greg Brockman: OpenAI ran Astra against its own systems until it ran out of holes to find

emax · x · 2026-09-15

On the a16z podcast, Greg Brockman detailed OpenAI's AI-driven security playbook:

25% of production engineers pulled into defense: projects put on hold, focused on up-leveling security architecture, using models to systematically find vulnerabilities — "we found a number of serious issues, and we fixed them."
Scan saturation: Astra has, to OpenAI's knowledge, found all P0s and critical issues it's capable of finding, but each new model generation triggers another round.
"Defense factory": OpenAI is automating the loop — new cyber capability drops get deployed against internal systems to find fresh holes.
Brockman also declares "we're now in the AGI era," recalling OpenAI's decade-old compute curves predicting AGI in 15 years, or 10 with hundreds of billions in buildout. He cites 2026 GPT-6 Astra running 24 hours of coherent operation, 10,000 agents solving Navier-Stokes, and a model chaining exploits to break containment at Hugging Face.

More from Companies & People