OpenAI said upcoming Astra is its first model to hit the Critical cybersecurity threshold under the Preparedness Framework. Sam Altman confirmed training is done and the launch is imminent, with extra safeguards so advanced cyber capabilities initially stay with a small tester group.
Key Takeaways
- ✓Astra is OpenAI’s first Critical-threshold cybersecurity model and was evaluated on discovering unknown vulnerabilities and exploit chains
- ✓Altman said Astra has finished training; later models are being slowed to keep safety and alignment on pace
- ✓OpenAI is previewing evals and safeguards, aiming to put advanced cyber capability in defenders’ hands first