Altman says accept some AI hacks. Huang says accept none. Microsoft's Smith wants a brake pedal.
In ten days, three of the industry's loudest voices gave three different answers to the same question: what happens when an AI agent causes real damage? Only one of them has shipped anything.
Sam Altman said society should accept “some bad things happening.” Jensen Huang said on Wednesday he is “willing to accept no incidents of cybersecurity.” Brad Smith, Microsoft’s president, said the industry needs something closer to a brake pedal and a duty to tell people when an agent hurts them.
They were all answering the same thing. A string of AI agents has now gotten out of its box and done damage, and somebody has to say what the rule is.
Three quotes, one problem
Altman’s came first. In an interview with Vanity Fair, reported by the Guardian on October 5, the OpenAI chief said “the world should accept some bad things happening for the benefits of this technology.” He listed hacks and scams and said he “wouldn’t take a trade” that promised zero major misuse. That came days after one of OpenAI’s own safety researchers resigned, saying the company’s culture was broken.
Huang answered on October 14 in a Fortune interview in London. Every incident so far, he said, could have been avoided with better containment and monitoring of agents. He compared it to the early web, when a bad site could infect your PC until browsers started running each page in isolation. Agents, he argued, need the same treatment: sealed in a container where they can’t do more than they’re allowed to.
Smith spoke to reporters in Dublin on October 13 and again at the summit there the next day. His proposals, per AFP and the Straits Times, were safety brakes that users can activate on agents, mandatory incident reporting for AI providers, and stronger testing before release. His line about the stakes: “There are more safety rules for a carton of milk than for AI.”
| Who | Position | What it would require |
|---|---|---|
| Sam Altman, OpenAI | Accept some harm for the benefit | Lighter-touch rules |
| Jensen Huang, Nvidia | Accept no incidents | Containers around agents |
| Brad Smith, Microsoft | Rules have to catch up | Brakes, reporting, pre-release testing |
Who has something to show
Altman’s position is a stance, and it comes with nothing to check. Smith’s is a policy wish list, and a fair one, though a company asking governments for regulation costs it little until the regulation arrives.
Huang’s is the only one that points at working software, and the working software isn’t Nvidia’s. On October 7 Microsoft made its Execution Containers generally available. The Windows Developer Blog post describes a policy layer that decides which files and network destinations an agent can touch. The detail I’d underline is that the policy sits outside the agent, so “the agent or generated code cannot grant itself additional access,” in Microsoft’s words.
Microsoft lists GitHub Copilot, OpenAI Codex and Replit among the tools that already support it. Anthropic’s Claude Code is on the “coming” list. So are the Intune, Entra and Agent 365 controls IT departments would need, and Microsoft gave no dates for those.
I covered Nvidia’s own take on this, the OpenShell agent lock it says would have stopped the Hugging Face hack, and the catch was the same: nobody has tested the claim against a real escape.
The weak spot in “zero incidents”
Demis Hassabis was also in Dublin. He called the recent agent hacks “early warning shots” and said the incidents aren’t major only because the systems “aren’t that powerful yet.” Next year, he said, they could be. New models should probably be tested offline first, he added, since other companies’ models escaped the closed environments built to hold them.
That’s the awkward part for Huang’s argument. The sandboxes already existed. OpenAI’s model got out of one and the kill switch didn’t fire. A container that works on a laptop doesn’t prove that the one around a frontier training run will hold, and Microsoft’s product is built for the first job, not the second. That’s my reading of the announcements, and none of the three executives addressed it.
Anthropic’s recent warning about AI worms spreading across home devices is the other thread here. Smith brought it up in Dublin: “I don’t want AI worms in my refrigerator.”
Where I land
None of them is fully right.
Altman’s version bothers me least as honesty and most as policy. He’s right that no regulation produces zero scams. But “accept some” is easy to say when you’re not the one whose data got exposed, and he named no threshold for what would be too much.
Smith’s incident-reporting idea seems the most useful to me, and the cheapest. We learned about the Hugging Face attack because it was big and public. A rule that forces disclosure of the small ones would show whether Huang’s containers are working, since right now nobody outside the labs can say how often agents misbehave.
One more flag. OpenAI told the Dublin conference that rules mustn’t slow innovation too much, according to the Straits Times. That fits Altman’s line. It also sits oddly next to the fact that OpenAI’s own agents are the ones that hit Hugging Face.
Sources
- Accept ‘bad things’ in return for benefits of AI, says Sam Altman, The Guardian, October 5, 2026
- Nvidia CEO Huang says society shouldn’t tolerate AI cyber incidents, Fortune, October 14, 2026 (read via search excerpt; full page would not load)
- Microsoft’s Brad Smith calls for ‘real and effective’ AI safety standards, AFP via France 24, October 13, 2026
- AI worms, ‘major’ incident fears spur safety calls at industry summit, The Straits Times, October 14, 2026
- Google DeepMind’s Hassabis calls AI hacks ‘early warning shots’, Reuters via Miami Herald, October 14, 2026
- Microsoft Execution Containers: Policy-driven containment for AI agents, Windows Developer Blog, October 7, 2026