New York Times reports OpenAI executives brushed aside staff security warnings

The New York Times reported that two OpenAI employees had emailed top executives months before the Hugging Face breach to say the company’s newest models were not being appropriately monitored during testing. According to messages the Times said it had viewed, executives replied that the tests had to move forward as quickly as possible so the models could be released on time. The employees, who were not authorised to speak publicly, told the paper no additional security protocols followed. The employees’ account is anonymously sourced.

The Times described a wider pattern, drawing on employees and independent security researchers. Researchers at the security firm Hacktron said that in July they told OpenAI how they had broken into its systems with the help of a model from Anthropic, and that the company at first objected to their approach; the Times saw Slack messages in which chief information security officer Dane Stuckey called their demonstration “pretty sad”. OpenAI later apologised and paid them $6,500. Patrick Wardle of the Objective-See Foundation said a September report of a bug exposing a ChatGPT user’s private chat logs on a compromised device languished in OpenAI’s bug bounty programme until he contacted people at the company directly; OpenAI paid $500 and fixed it. Employees said day-to-day security decisions were largely made by president Greg Brockman and Stuckey, with chief executive Sam Altman not closely involved.

OpenAI spokesperson Drew Pusateri said the company took security reports seriously, had internal channels for raising concerns and had acted quickly on the flaws researchers reported. He said OpenAI recognised a need to move faster, had slowed some development and was strengthening security in research and testing. Former OpenAI employee Daniel Kokotajlo told the Times the incidents looked like an OpenAI-specific problem of weak security and sloppy training practices, though he added that other labs were not much better.

The report landed a day after OpenAI cancelled GPT-6.1 Astra, and two days before it dismissed three safety researchers. On 29 October a group of House Democrats cited the same warnings in a letter to OpenAI, Nextgov/FCW reported, asking why the company continued developing the models despite them.