Altman: AI models will expose many more serious software vulnerabilities than expected, after OpenAI's 'PHP week'

OpenAI CEO Sam Altman says AI models are improving quickly at finding software vulnerabilities and that the industry has barely started pointing them at the code that exists. In an interview with Stripe co-founder John Collison on the Cheeky Pint podcast, published October 29, 2026, Altman said he expects "a world of many more serious vulnerabilities than we expect."

PHP week and the vulnerability problem

Collison opened with "PHP week," the episode's name for the period in which models uncovered critical PHP vulnerabilities. Altman gave two reasons the bugs surfaced when they did. First, the models "continue to get better fairly rapidly." Second, people had not really used them "to go look at everything out there," given the amount of surface area and code on the internet.

Collison argued that patching software is not enough and that architectures may need to change, which is harder but more durable. Altman agreed: "That seems to me what needs to happen." He said the task looks very hard, "except we have this amazing new tool to do it with." He pointed to a new generation of startups building new application architectures that assume models keep getting smarter. Some, he said, plan to build a whole new stack, including re-architecting what a router, an operating system or an application is. He said that would have seemed insane two years ago.

Cyber capabilities and iterative deployment

Collison asked whether training models for cyber capabilities was a mistake. Altman said the criticism has merit: people can fairly ask who requested these capabilities. He said the AI field has not explained well why the potential downsides are worth it, and described the current pitch (cures and cheaper services, alongside lost jobs and a surveillance state) as "a terrible sales pitch."

Altman said he is most proud that OpenAI stuck with iterative deployment, putting tools in people's hands, even when the policy was "hugely unpopular in the AI safety community." He called it "the only kind of good path."

Other topics in the episode

According to the episode description, the conversation also covers:

• Why sandboxing superhuman models is an unsolved safety dilemma (timestamp 00:08:54)
• OpenAI's internal experiments on complex math problems such as Navier-Stokes (00:15:47)
• Why test-time compute unlocks large agent swarms (00:23:12)
• How robotics will scale future gigawatt data centers, and revenue per gigawatt (00:25:17 and 00:37:53)
• Why businesses must adapt as personal agents replace traditional web forms and mobile apps
• A segment titled "Token pricing is fake" (00:49:27)

Codex outage anecdote

Altman also described an incident at OpenAI in which Codex went down during a SEV, and the reaction inside the company was a sense of "Whoa, what if you couldn't?" write code. He used the story to illustrate how quickly people have come to expect software to be built by computers on request, particularly children, who he said now build full 3D video games in minutes.

What this means

Altman's comments frame AI-driven vulnerability discovery as a coverage problem as much as a capability problem. The models are already good, and the number of findings depends on how much code people point them at. His stated conclusion, shared with Collison, is that patching individual bugs is insufficient and that architectural redesign is the more durable response. He also ties that redesign to the same models, calling them the tool that makes a rewrite feasible. These are Altman's views, not a measured result. The episode description lists his positions on sandboxing, agent swarms and compute scaling, but the details behind those items are in the full conversation.