GPT-6 Cyber stays in bounds more often, and agent security gets pricey
Last Friday, OpenAI explained how it watches its own agents, after reports that some broke out of a test this summer. Now it launches GPT-6 Cyber, its new cybersecurity model. It finds weak spots in software and helps write fixes. OpenAI says it is trained to stay inside its "authorized scope", meaning only the systems it is allowed to test.
In OpenAI's testing, GPT-6 Cyber went out of bounds less than 1% of the time, down from a higher rate before, The Verge reports. TechCrunch asks the obvious question in its headline: is that enough? AI labs face more scrutiny about how they protect digital infrastructure.
The money follows the worry. TechCrunch reports that Palo Alto Networks is in talks to buy Clutch, a startup that secures AI agents, for $500 million. Clutch was last valued at $85 million, according to a source. These are talks, not a done deal.
Less than 1% sounds small. But a hacking agent can run thousands of times, and each slip touches someone else's systems. These numbers come from OpenAI's own tests, so outside checks matter.
