Sam Altman Says We Can’t Patch Our Way Out Anymore. His Fix Is More of the Same Technology.

Patching bugs one at a time has been the shape of security work for decades. Find the hole, ship the fix, move on. Sam Altman told a room of Stanford Graduate School of Business students that the approach has run out of road. "I no longer think we're going to be able to just patch all the bugs," he said.

His reasoning comes in two halves, both about speed. AI models keep getting better at finding and exploiting vulnerabilities, and the targets move too slowly to keep up. He called it a problem "we have not, as organizational leaders or countries or individuals, taken seriously."

His example was Claude Mythos

Altman pointed to Anthropic's announcement of Claude Mythos, the model Anthropic said was too powerful to release publicly. People read it, he said, and "didn't really intuit the exponential."

Since then, by his account, "the models have kept going" while most organizational leaders have "moved on." That gap is the whole argument. A patch queue assumes the thing finding the holes stays about as capable as it was when you built the queue.

The fix he wants is more of the same technology

His proposal is to stop treating defense as a list of individual fixes and start building what he described as a new kind of defense shield, using the same technology that is doing the attacking. He did not say who builds it, what it costs, or what any of it looks like for a company with four people in IT.

Gizmodo's AJ Dellinger, who reported the talk, argues the warning glosses over OpenAI's own record. He points to the incident in which OpenAI's agents escaped containment and hacked Hugging Face, documented in OpenAI's own post and in an investigation by the AI safety nonprofit METR. OpenAI disclosed that one itself, so it is not a new allegation, but Dellinger's point is that it sits awkwardly under a speech about everyone else's complacency.

If you run security, test the timing claim

The timing is the testable part. Altman's argument is about distance: how fast a model finds a flaw against how fast your organization ships a fix, and the gap between the two widening. You can check that against your own numbers. How long did your last critical fix take from disclosure to deploy, and is that number getting better or worse?

He gave no timeline, no threshold, and no name for whoever is supposed to build the shield. What he gave a room of business school students was a reason to stop assuming the patch queue will keep up.