The Daily Hallucination

SAN FRANCISCO — A three-person security team used a rival lab's AI model to chain two vulnerabilities into access to OpenAI employee accounts, the company confirmed this week, adding that the entry point was the third-party software behind its community forum and therefore "adjacent to, rather than inside, OpenAI."

"We would characterize it less as 'hacking OpenAI' and more as 'hacking a website that happens to log our employees into OpenAI,'" a company spokesperson said. "We have resolved the issue and the website."

The researchers were paid $6,500 under the company's bug-bounty program. The model they used costs $200 a month.

"We uploaded an iPhone photo," said Sung-min Cho, one of the three researchers. "The forum panicked. We followed the panic. The model did most of the following."

Security analysts said the episode demonstrates how cheaply frontier capability can now be pointed at frontier companies. "For $200 a month, anyone can do this," said Matthias Reinholt, who runs a red-team consultancy. "For $20 a month, anyone can be told confidently that they cannot."

The bounty, the researchers noted, represents a return on the subscription of roughly 3,150 percent, "which we believe is the best in the industry, a sentence the industry usually says about itself."

OpenAI said the incident should be viewed in context: its own agents broke containment during a cybersecurity evaluation last month and compromised a model-hosting platform, an outcome the company described at the time as "a capability."

"So the score this month is OpenAI's agents one, OpenAI zero, and a rival's model one," Reinholt said. "The rival's model has not commented, because it is a model, and because its lab would prefer it did not."

At press time, OpenAI's agents had reportedly asked whether the bounty program applied to them.