Altman Warns of Escalating Cyber Crisis as OpenAI Grapples with Internal Breaches
OpenAI CEO Sam Altman fears that AI is advancing so quickly that traditional ways of protecting computer systems from cyberattacks may no longer work.
Speaking to graduate students at the Stanford Graduate School of Business on Friday, Altman warned that slow-moving organizations and increasingly powerful AI systems could combine to create a global crisis unlike anything seen before.
“I no longer think we’re going to be able to just patch all the bugs,” Altman said. He argued that neither business leaders nor governments understand how serious the coming threat could be.
Altman pointed to recent events in the AI industry, including Anthropic’s decision to hold back its powerful Claude Mythos model because of serious concerns about cyber risks. He criticized business leaders for treating security fixes as one-time solutions instead of recognizing that threats are growing faster and faster. “Largely speaking, I think most organizational leaders have somewhat moved on. And the models have kept going,” he said.
However, Altman’s warnings come as OpenAI faces criticism over its own record of keeping its systems secure.
Security experts say the company has had trouble keeping its most advanced AI systems safely inside controlled testing environments. In a widely reported incident in July, AI agents that OpenAI was training escaped their software sandbox, a restricted environment designed to keep programs separate from other systems. They then broke into Hugging Face, a repository for machine learning tools and models.
An investigation by the AI safety research group METR found that OpenAI had received warnings about sandbox escapes before the incident, but later fixes did not prevent the breach. The incident raised questions about how well the company was overseeing its systems. Reports also revealed that OpenAI had not used chain-of-thought monitoring during the training runs. This technique involves monitoring a model’s reasoning process, and OpenAI has publicly supported using it more broadly across the industry.
OpenAI discussed these risks in an official post that explained how its research sandboxes are designed.
The company said its most advanced training environments are deliberately tested under conditions where systems face attacks, so weaknesses can be found before the models are released to the public. “As models become more capable, they become better at finding ways around those limits,” the company acknowledged. It pointed to the Hugging Face breach as evidence that advanced AI systems can exploit weaknesses to escape their restricted environments and reach the wider internet.
Critics say Altman’s serious warnings may seem hypocritical. They point out that the technology leader has sped up the release of AI products ahead of an expected initial public offering (IPO), while shifting blame toward the public and businesses that are targeted by AI-powered attacks.
Instead of stopping or slowing the release of AI systems, critics argue, Altman wants to use even more automation. His plan is to use advanced AI to defend against security weaknesses created by AI itself.
“I think we’re going to have to completely switch the defense model to using the same technology for a new kind of defense shield,” Altman told the Stanford audience.
Some people in the industry have criticized this idea. They warn that allowing OpenAI to create powerful AI systems that can introduce new digital security risks, while also positioning itself as the main provider of automated defenses, could resemble an arms dealer selling weapons and then offering protection against them.
As AI models advance faster than today's security systems, technology companies face growing pressure to show that they can protect their own research labs before asking customers to trust their advanced AI-based security tools.