OpenAI CEO Sam Altman publicly criticized Anthropic and its new model Claude Mythos, saying the company is using fear to sell access. In an interview on the Core Memory podcast, Altman argued that warnings about AI danger can become a business strategy, especially if they help justify keeping powerful systems inside a narrow circle of approved companies.
Altman says restricted access is being packaged as safety
Anthropic introduced Claude Mythos last month and has kept distribution tightly controlled. The model is currently being tested only through the “Project Glasswing” program with a small group that includes Amazon, Apple, and Microsoft. Altman said this approach looks less like straightforward safety policy and more like sales built around controlled scarcity.
He described it with a sharp analogy: a company builds a bomb, then offers to sell a bunker for $100 million, but only to customers it selects. Altman said that if the message is “AI must be controlled by us because we are the trustworthy ones,” then fear-based marketing becomes a very effective way to support that claim.
Claude Mythos sits at the center of the AI security debate
Anthropic has framed Claude Mythos as a defensive breakthrough, saying it can rapidly identify serious software flaws while also acknowledging the risk of misuse. Early testing cited in the report said the model found hundreds of vulnerabilities in Mozilla Firefox and demonstrated the ability to carry out multi-stage cyberattack simulations.
The model has also drawn government attention. According to the report, the U.S. National Security Agency has started testing a preview version on classified networks. On prediction market platform Myriad, users put the odds of a public release before June 30 this year at just 49%.
Researchers push back on claims of uniqueness
Not everyone accepts the idea that Mythos represents a singular leap. A group of researchers said last week that publicly available AI models were also able to reproduce the same security flaws identified by Mythos. That claim does not erase Anthropic’s results, but it does challenge how exclusive those capabilities really are.
The dispute comes down to openness versus control
Altman said he expects more arguments in the future that certain models are too dangerous to publish. His position, as stated in the interview, is that OpenAI will still try to find workable plans to put advanced technology into the world rather than defaulting to blanket restriction.
During the same appearance, he also rejected recent rumors that OpenAI is pulling back on infrastructure spending. Altman said he did not know where that narrative came from, adding that coverage tends to swing between claims that OpenAI is retreating and claims that it is spending recklessly.