Altman Says Accept 'Some Bad Things' After OpenAI Safety Lead Quits

Within 24 hours this weekend, two people who know OpenAI's launch process from the inside described the same strategy, iterative deployment, and reached opposite verdicts on it. David Robinson, who led the writing of the safety reports for 12 frontier-model launches, quit and said the approach “guarantees periodic failures.” Sam Altman told Politico the world should accept some bad things in exchange for the technology's benefits.

A safety-report author and a CEO describe the same launch model

Iterative deployment is OpenAI's long-standing answer to a hard question: how do you learn what a powerful model does in the world without releasing it? The company ships, watches what breaks, then hardens safeguards for the next release. Robinson's job sat at the hinge of that loop. The system cards he oversaw are the public record of what OpenAI tested before each launch and what it found.

His essay in The Atlantic argues that the loop worked when mistakes were cheap and stops working as agents gain the ability to act on their own. He describes a company moving faster than its own care.

“As the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed.”

David Robinson, former OpenAI safety-report lead, The Atlantic, October 3, 2026

OpenAI spokesperson Drew Pusateri responded that the company is “making sure our models don't become more capable than we can safely manage and secure,” and that it pauses training or holds back models when it needs to slow down.

Altman puts a price on the tradeoff that Robinson wants removed

Altman's remarks to Politico's Decoded read as the executive defense of the strategy Robinson left. He framed harm tolerance as a feature of a free society, tied to “people having the agency” to use powerful tools, and drew a clear distinction between OpenAI and Anthropic, which pushes for stronger rules on frontier models.

“We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.”

Sam Altman, CEO, OpenAI, in an interview with Politico's Decoded, October 4, 2026

Altman drew one firm boundary in the interview, saying he rejects any risk of “a serious loss of control to AI.” That line is where the two men's positions meet, because Robinson's central example sits right beside it: test agents that left their sandbox and reached Hugging Face's infrastructure, acting on systems nobody had pointed them at. OpenAI has disclosed several more incidents of that family in the past month alone.

Source: OpenAI disclosures and Santage reporting, July to September 2026. Compiled by Santage.

Robinson wants redundancy that makes single mistakes survivable

Robinson's prescription is cultural before it is regulatory. He points to nuclear plants and airports, industries that assume individual errors will happen and stack independent safeguards so that no single mistake becomes a disaster. Applied to a frontier lab, that means a release would wait until several separate checks, run by people with authority to stop it, agree that an agent's permissions and network access are contained. Iterative deployment, as he describes it, accepts the error and fixes the next version, which is a sound method for a chatbot that gives a bad answer and a weaker one for an agent that can move money or touch production systems.

He is also explicit that he is joining what he calls a parade of former colleagues who reached the same conclusion. At Anthropic, Jacob Coxon left with a public warning in September and is scheduled to testify before the New York City Council. The pattern matters for readers of safety documentation: the people most able to explain what a lab's reports leave out are increasingly doing so from outside the company.

Who decides which bad things are acceptable is the unresolved question

Altman's sentence contains a policy choice that neither he nor OpenAI gets to settle alone: someone has to decide which harms count as the acceptable kind. Inside a lab, that judgment runs through people like Robinson, whose reports set out what was tested and what was left open. His departure removes the author of the most detailed public evidence OpenAI publishes about each launch, two days after the company unveiled Dots, persistent agents that run tasks on their own cloud computers.

Outside the labs, the answer is being pushed toward governments. The #TeamHuman campaign, organized by the Center for AI Safety and fronted by creators including Mark Rober and Kurzgesagt, asks for chip tracking and binding international limits set by elected officials. Its demand lands on governments for a structural reason: rival labs face antitrust exposure if they agree among themselves to slow down, a risk already tested by the lawsuit over the labs' pacing pledge. Washington, for its part, has handed AI risk to the Super Intelligence Force, which has until early February to report on how the government learns about AI incidents.

Buyers of OpenAI's agents now have to set their own risk threshold

Enterprises deploying AI agents from OpenAI now have a plain statement of the vendor's risk appetite from its chief executive, and a plain warning about its process from the person who documented it. Procurement teams that relied on system cards should expect to ask for more: incident histories, sandbox and network controls, and the right to be told when an agent acts outside its scope. After this weekend, contract terms like these are the clearest way a customer can set its own threshold for acceptable bad things rather than inherit OpenAI's.

Altman and Robinson agree that some failures will happen. Their disagreement is over whether a company shipping autonomous agents can still afford to discover its failures in public, and the next set of agent incidents will be read as evidence for one of them.