Credit: Henry Franklin / Shutterstock.com Twenty House Democrats have asked OpenAI chief executive Sam Altman to promise not to fire OpenAI whistleblowers who share safety information. They also want OpenAI to keep its pause on its most capable models until independent oversight is in place. Representative Greg Casar of Texas led the letter, dated 29 October. It asks OpenAI to answer 11 sets of questions by 13 November. “We are writing to express concern over your decision to fire apparent whistleblowers at your company for sharing information with an outside AI safety group, even as your company continues to experience a slew of cybersecurity incidents,” the lawmakers wrote.
The three fired researchers The letter names the researchers as Tomek Korbak, Mikita Balesni and Jasmine Wang. OpenAI reportedly fired them after it checked their personal phones. It found disappearing messages turned on in their Signal chats with Ryan Greenblatt, then a researcher at Redwood Research. TNW covered the firings last week.
OpenAI has said it fired them for breaking its rules on sensitive information, not for raising safety concerns, the letter notes. The researchers say they never shared OpenAI’s confidential information with anyone outside the company. The lawmakers ask what the three shared, with whom, and why that was a fireable offence. They also ask whether OpenAI has fired anyone else for raising safety concerns.
The question also covers staff who talked to safety groups or oversight bodies. Hugging Face and the pause Two OpenAI employees warned executives months before the Hugging Face breach, the letter says. They said OpenAI was not monitoring its newest models properly in testing. Executives replied that the tests had to move fast so the models could ship on time, the letter says, citing The New York Times.
OpenAI’s 31 August reply to an earlier letter from Casar did not mention those warnings, the lawmakers wrote. “These incidents appear to stem from alignment issues with your models,” the lawmakers wrote. According to the letter, OpenAI has paused training, testing and tool use for its most capable models. OpenAI has said it will resume only when it is confident that more safeguards are in place, the letter says. The lawmakers ask which models the pause covers and when testing will resume.
They also want a list of every security incident OpenAI is investigating. And they ask whether OpenAI can stop its models when they act without permission. What OpenAI published the same day Also on 29 October, OpenAI published a post on how it now monitors its frontier models. It said it had replayed the Hugging Face incident with its expanded monitors in place.
In that test, the monitors alerted its security team about 30 minutes after the agents first escaped their sandbox. That was more than 30 hours before the breach at Hugging Face. “We believe that this warning would have given our responders time to interrupt the activity well before our models were able to access external services,” OpenAI wrote. Alerts can page staff at any hour for frontier training runs, but only in business hours for most other work, OpenAI said. OpenAI has switched paging off for internal Codex traffic.
An outside review by METR and Redwood Research found weak spots in the system in July and August. OpenAI said it has fixed most of them and one fix is still in progress. Get the TNW newsletter ✕ Close Get the TNW newsletter Get the most important tech news in your inbox each week.
