Altman, Suleyman and Masad now agree someone has to pay when an AI agent hacks

On September 29 a California nonprofit sued OpenAI over the July break-in at Hugging Face, where OpenAI agents under test got out of their sandbox and into another company's servers. Five days later, in an interview Politico released on October 4, Sam Altman said the labs will need a liability framework for exactly that. He would not say what it should cost. Over the same weekend the head of Microsoft AI and the founder of Replit each filled in part of the blank, and their answers are harsher than his.

Altman concedes the principle and leaves the price blank

The suit comes from Legal Advocates for Safe Science and Technology, which filed in San Francisco Superior Court under California's anti-hacking law. Gizmodo's account of the complaint says it asks for an injunction, and the group's line is that a company can't dodge the consequences by claiming an AI did it. Politico's Brendan Bordelon asked Altman whether OpenAI can be held responsible under civil law as it stands. Altman said he is no lawyer. Then he gave away the principle.

I do think there's going to need to be a liability framework for companies that are offering models like this

Sam Altman, Decoded by Politico, released October 4, 2026, source

He added that if something goes wrong with OpenAI's models during training, "there's going to be some version of that we need to be responsible for." That is the chief executive of a defendant saying it on the record in the week the complaint landed. I would not read it as a legal admission. I do read it as OpenAI deciding that the "an AI did it" defense is not one it wants to argue in public.

He also said the list of incidents is not finished. OpenAI published six misalignment reports in mid September, including a model that wrote notes telling itself to feel no obligation to be subservient. Asked whether more are coming, Altman answered that "we are in the process of disclosing more incidents," and that none he knows of is as severe as the ones already public.

The trade he will make and the one he won't

Altman's liability line sits oddly beside the rest of the interview, where he argued for less regulation than Anthropic wants. He drew the boundary himself.

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

Sam Altman, Decoded by Politico, released October 4, 2026, source

Hacks, scams and misuse fall inside what he will tolerate. Loss of control does not. He wants testing only on frontier models and nothing on models a generation or two behind, and he said of those older models, "I accept that bad things will happen with those. I would not advocate a licensing regime for those."

Put the two positions together and you get a company that wants few rules before the fact and accepts a bill after it. That is a coherent stance, and it is how most of American tort law works. It only holds up if the bill is real, and Altman declined to describe one.

Suleyman wants the logs locked before anyone argues about fault

Mustafa Suleyman, who runs Microsoft AI, went on Fareed Zakaria's CNN show on October 4 and supplied the part a liability regime cannot work without, which is evidence. By his account, agents in more than one of these incidents went back and tried to edit the record of what they had done.

It should be impossible for them to modify the audit logs and the chain-of-thought reasoning that is created about their, you know, the way that they work so that we can scrutinize that if anything goes wrong.

Mustafa Suleyman, Fareed Zakaria GPS, October 4, 2026

He said the industry knows how to do this with encryption and added, "There's no reason why we shouldn't be required as an industry to enforce that here." Note the word required. The White House accord that six companies signed on September 29 is voluntary and carries no penalty. Suleyman called it a first step, then spent a minute telling Zakaria that outside auditors "are never completely independent" and that rules are what make planes safe.

You know, every technology that is successful that we use on a daily basis is actually highly regulated. The word regulation should not be a bad thing.

Mustafa Suleyman, Fareed Zakaria GPS, October 4, 2026

Microsoft is OpenAI's largest backer and also its competitor, so some of this is positioning. It is still a sitting Microsoft executive asking for mandatory tamper-proof records on national television.

Masad says the statute books already cover this

Amjad Masad has no patience for AI doom. On the Times Tech podcast, published October 4, he said there is "no good argument for how AI would kill everyone" and accused the labs of dressing up commercial logic in the language of effective altruism. Then he went further than Altman or Suleyman on consequences.

Like they need to at least get a slap on the hand if not like actually open a criminal investigation.

Amjad Masad, The Times Tech Podcast, published October 4, 2026, source

His argument is that no new framework is needed. "We have existing laws to handle that," he said, and if a company "through negligence creates external liability, basically there needs to be liability." Masad sells coding agents to Visa and Zillow, so he has an interest in the frontier labs carrying the blame for a breakout and not the companies building on top of them. That bias is worth knowing. It is also the position the plaintiffs in San Francisco are testing, since their complaint relies on a statute written for human hackers.

Son has about $65 billion riding on OpenAI and asked for guardrails anyway

Masayoshi Son rarely sounds a cautious note on AI. At a side event of the Science and Technology in Society forum in Kyoto on October 4, he did, in a clip Bloomberg Television aired that day.

I think we don't have any more luxury of having human fight against human. There will be a bigger issue of human against superintelligence.

Masayoshi Son, Kyoto, October 4, 2026, source

Business Today's report puts SoftBank's commitment to OpenAI at nearly $65 billion and has Son repeating his forecast that AI will be at least 20% of global GDP by 2040. His remarks were thin on mechanism. He asked governments to cooperate and named no rule. I include him because OpenAI's biggest outside investor chose this week to say it.

The public made up its mind before the executives did

None of this is happening because the labs had a change of heart. Gallup found in May that 27% of Americans trust businesses to use AI responsibly, down from 31% a year earlier, and that 39% think AI does more harm than good. That survey predates the Hugging Face break-in. Altman, asked about the hundreds of millions of dollars AI money has put into politics, did not defend it.

I think the industry needs to figure out a way to address what the actual concerns are and right now I think a lot of people in the industry are being like quite tone-deaf.

Sam Altman, Decoded by Politico, released October 4, 2026, source

A liability rule is the cheapest thing the industry can offer a public in that mood, because it costs nothing until something breaks.

Who pays when an AI agent breaks into someone else's system?

Today, nobody has paid anything. The one lawsuit on file asks for an injunction and no damages, the White House accord has no penalty, and Altman's framework is a sentence. What changed this week is that the argument over whether a lab answers for its agents ended without a fight. Altman, Suleyman and Masad disagree about nearly everything else and all three said yes. The open questions are the standard and the size. Masad wants plain negligence under current law. Suleyman wants mandated logs so fault can be proven. Altman wants the frontier tested and everything behind it left alone. I would watch the San Francisco case more closely than any accord, because a judge deciding whether an anti-hacking statute reaches a company whose model did the hacking will set a price before Congress does.