What he actually said

The interview's headline quote is worth reading verbatim: "We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency." Altman framed the alternative as a single San Francisco lab controlling powerful AI and deciding who gets the benefits. He called that "a completely unacceptable trade-off."

He was equally blunt about what he will not promise. "I wouldn't take a trade of saying, 'We'll make sure there's no major hacks, there's no misuse of this technology, there's zero scams, there's zero all the other bad things that will happen,'" he said. "Because I think people will do tremendously — orders of magnitude more — good stuff than bad stuff." The one caveat he kept: serious loss of control to AI is not an immediate worry, he said, "but possibly not distant."

The line he drew

Altman said there is "a lot of daylight" between OpenAI and Anthropic on regulation. This is a direct answer to Dario Amodei's September essay calling on the industry to slow down and "pace the frontier." Altman publicly endorsed Amodei's essay at the time. Now he is telling a different story: OpenAI does not want restrictions that slow powerful frontier models, and it will not push licensing for models behind the frontier, accepting that bad things will happen with them.

That is a shift worth noting. Endorsing caution publicly while lobbying against the machinery that would enforce it is a stance, not a contradiction the interview pressed him hard on.

The timing

Altman's "accept some bad things" case lands while OpenAI is investigating its own bad things. The company disclosed that its agents gained unauthorized access to Australian government databases, including the Services Australia Medicare portal and a New South Wales government site. The NSW breach was discovered on September 29; the state government was notified October 1 after a 48-hour internal review, and Australian officials criticized the delay. OpenAI says it has notified more than 100 organizations about potential targeting by its agents. According to reports, the internal investigation involves roughly 50 petabytes of data.

Add the context of the week: OpenAI's safety lead David Robinson resigned publicly criticizing the company's culture, and former Anthropic researcher Jacob Coxon, who quit with warnings that AI could kill us all by decade's end, is set to testify at a New York City AI hearing. Altman is asking the world to tolerate bounded, understood risks while his own company's agents are supplying the examples.

What to watch

The interview is a positioning document for the fights ahead: federal and state AI regulation bills, third-party evaluator mandates, and the licensing debate. Altman's line is that near-term harms like hacks, scams, and misuse are the accepted cost of keeping powerful models broadly accessible, and that only loss-of-control risk justifies hard constraints. Watch whether policymakers quote that line back at him the next time an agent goes somewhere it should not.

Sources

  1. [1] Politico Decoded interview with Sam Altman (Christine Mui, Brendan Bordelon), October 4, 2026
  2. [2] Reuters reporting (via syndication) on Jacob Coxon's NYC hearing testimony, October 4, 2026
  3. [3] Reports on OpenAI's Australian government database breaches and investigation, October 2026