openai built an AI that might be too good at hacking. so it's not shipping yet

Astra is the first OpenAI model ever flagged as possibly 'Critical' for cyber capability — and the company told the White House before it told the… Aliteq Lena Fischer · AI & Local Compute Editor

OpenAI told Axios on August 7, 2026 that internal testing of its unreleased Astra model couldn't rule out 'Critical' cyber capabilities under its own Preparedness Framework.

'Critical' is OpenAI's highest cyber-risk tier: a model that can independently find and build working exploits for unknown vulnerabilities, or plan and run a full cyberattack from just a broad goal,…

OpenAI is pausing internal work on Astra that doesn't meet upgraded security requirements, and adding isolated testing environments, tighter network restrictions, stronger weight encryption, and…

OpenAI says it voluntarily briefed the White House on the delay before going public with it.

This is reportedly the first time a frontier AI lab has deliberately slowed a model's development specifically over cyber-offense capability, rather than reacting after an incident.

I think this is OpenAI doing something genuinely rare: pausing publicly and on its own terms, before an incident forces the conversation instead of after. That's worth crediting. But it lands in the…

Read the full story

Aliteq openai built an AI that might be too good at hacking. so it's not shipping yet Read the full story on Aliteq