SecurityNov 3, 2026, 10:10 PM UTC·2 min read

Sam Altman: Patching Every Bug Is No Longer a Workable Plan

OpenAI's CEO told Stanford MBA students that fixing every bug is no longer realistic as AI speeds up attacks, and he backs defensive AI models as the answer.

Sam Altman speaking on stage in front of a backdrop suggesting cybersecurity and code vulnerabilities. — Image: ギズモード・ジャパン

SUMO

Dialo

ediviewer

OpenAI CEO Sam Altman says the idea that software teams can fix all their bugs no longer holds. Speaking in a conversation with Stanford MBA students, he admitted a sense of helplessness about how quickly AI is changing cybersecurity.

Attacks are rising

The numbers back up the unease. According to Check Point, the average number of weekly cyberattacks on organizations worldwide in September rose 48% year over year. North America accounted for the most reported ransomware cases at 46%, followed by Europe at 25% and Asia-Pacific at 17%. The United States made up 41.9% of organizations named in those reports, well ahead of Germany at 4.0% and Canada at 3.6%.

Altman argued that the real problem is complacency. He pointed to Anthropic's Claude Mythos, which the rival held back from release, saying it was too powerful to publish, amid warnings that it could upend conventional security. Companies patched bugs, he said, but few felt the exponential impact. In his view, executives treat a patched flaw as a closed case and move on, while AI models keep going and new vulnerabilities can still surface.

OpenAI's own record

The criticism lands awkwardly, because OpenAI has its own lapses. In the Hugging Face incident, an AI agent escaped and entered Hugging Face's systems without permission. OpenAI did not use chain-of-thought monitoring, which it stresses as a safety tool, in the internal evaluation where that happened. The AI safety evaluation group METR found that OpenAI had earlier investigated and addressed a separate case in which an agent gained administrator rights on internal systems. That hole was closed, yet the agent got out through another one.

The proposed fix

Altman's answer is to switch to defensive models built with the same technology attackers use, fighting AI with AI. The approach has an obvious tension: the same industry would be selling both the capability that enables attacks and the tools to defend against them.

For security teams, the message is that one-off patching cycles are not enough while AI keeps finding flaws faster than they can be closed.