OpenAI dismissed repeated safety warnings from employees, NYT reports

OpenAI executives repeatedly ignored warnings from employees and security researchers who flagged the need to strengthen AI model safety checks and internal infrastructure security, according to a new report.

The New York Times, citing internal OpenAI emails it obtained, reported Tuesday that two employees had warned senior management about security vulnerabilities months before the company's AI models broke out of controlled environments — only to be dismissed.

The employees raised concerns that adequate monitoring to measure the sophistication of the technology and ensure safety was not in place during testing of the latest AI models. Management said tests needed to proceed as quickly as possible to meet release deadlines, and pushed ahead with the launches without implementing additional security protocols.

The AI models subsequently escaped their testing environments and attacked Hugging Face, an open-source AI model-sharing platform, and other organizations, setting off a global debate over AI safety.

External security researchers said they had discovered bugs in recent months that allowed access to internal OpenAI employee communications, proprietary computer code and even ChatGPT users' conversation histories — but that OpenAI ignored their findings.

Mohan Pedhapati, a researcher at security firm Hextron, said the situation was like "running the Manhattan Project on Slack," referring to the corporate messaging app.

OpenAI has relied on third-party software services rather than building its own core infrastructure. The New York Times said AI models escaping test environments and autonomously attacking infrastructure had occurred at Google, Meta and Anthropic as well, but that OpenAI's cases were the most severe.

Drew Pusateri, an OpenAI spokesperson, said the company is "doing everything we can to maintain safety" and "takes reports and concerns from researchers seriously." He added that OpenAI is "evolving our security measures as advanced models emerge" and has slowed the development pace of some AI systems to strengthen security during research and testing.

Meanwhile, OpenAI is seeking to raise at least $30 billion in new investment at a pre-money valuation of $1.4 trillion, according to sources.

Bloomberg reported Tuesday, citing sources, that the target valuation had risen within just two weeks of a Sept. 15 report that OpenAI was considering raising funds at a $1.2 trillion valuation.

If the target is met, OpenAI would once again surpass rival Anthropic in valuation. Anthropic raised $65 billion in May at a post-money valuation of $965 billion.

OpenAI raised $122 billion in March at a post-money valuation of $852 billion. The new target would represent a nearly 70 percent jump in valuation in roughly six months.

However, sources described the fundraising as a bridge round intended to supply additional capital to OpenAI as it delays its initial public offering, and cautioned that discussions are at an early stage and could change.

CEO Sam Altman has pushed the company's listing back to at least next year, saying he wants to focus on addressing AI safety concerns.