OpenAI Locks Down Astra Over Cyber Capability

Astra Crosses a Line OpenAI Drew Itself

OpenAI has concluded that one of its upcoming models is capable enough to need extra safety measures before launch. The model, called Astra, can find more security vulnerabilities than the most advanced OpenAI model currently available to the public, company officials told reporters on a conference call on 1 September. It also does that work using less computing power.

Key Takeaways

• Astra is the first OpenAI model to trigger the tougher safeguards required by the company’s safety protocol, a threshold that had been theoretical until now.
• Access to Astra’s most advanced cybersecurity capabilities will be limited to a small group, with no release date given beyond “soon.”
• OpenAI restarted its largest model training run on 28 August but is still holding back some smaller experiments.

What the Model Can Do

The capability that set off the internal alarm is autonomy. “With the right tools and access, Astra can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step,” said Amelia Glaese, an OpenAI vice-president overseeing its safety work.

Under OpenAI’s safety protocol, additional guardrails are required when a model shows two abilities together: spotting and using new cybersecurity vulnerabilities, and planning and carrying out a detailed, original attack strategy, both with minimal or no human involvement. Astra is the first to meet that bar.