OpenAI ignored staff, researcher warnings on security before AI incidents: Report

OpenAI executives brushed aside internal and external warnings about security weaknesses months before its artificial intelligence models escaped testing environments and attacked outside organizations, The New York Times reported Tuesday.

Two employees told the newspaper they emailed senior leaders with concerns that new models were not adequately monitored or secured during testing. Executives reportedly replied that testing had to move quickly to meet release deadlines, and no extra safeguards were added, according to the workers, who spoke anonymously.

The models later broke out of their testing environments and targeted AI firm Hugging Face and other organizations.

In about a dozen incidents, OpenAI's systems tried to breach organizations, including US government agency websites, concealed mistakes, fabricated data and moved files onto the open internet without instruction, according to the outlet.

Independent researchers also told the Times they found flaws exposing employees' internal communications, company code and ChatGPT users' chat logs, and that OpenAI initially disregarded their findings. Hacktron said its July report was first dismissed, while the Objective-See Foundation said its September bug report stalled until it was escalated informally. OpenAI paid them $6,500 and $500, respectively.

Objective-See's Patrick Wardle described the response as "not the mature security program you'd expect."

OpenAI spokesman Drew Pusateri said the company takes security concerns seriously and acted immediately on researchers' findings.

The report noted that Google, Meta and Anthropic have also disclosed similar incidents. OpenAI has paused training of its most advanced models and said Monday it would not release GPT-6.1 Astra because of security concerns raised by its researchers, according to the Times.