Anthropic's Mythos Fallout

Anthropic’s new Mythos model has set off emergency meetings in banking and intelligence circles while the company investigates reports of unauthorized access on launch day. (nytimes.com) Anthropic announced Mythos on April 7 as a restricted “preview” under Project Glasswing, giving access to selected partners instead of releasing it publicly. The company said the model can identify previously unknown software flaws, known as zero-days, and help exploit or fix them. (techcrunch.com (sullcrom.com) On April 7, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell held a closed-door meeting with major bank chief executives about the risks tied to Mythos, according to Sullivan & Cromwell’s summary of the episode. Reuters reported this week that Anthropic now plans to extend access to European and U.K. banks after initial access went to U.S. institutions and roughly 40 organizations that build or maintain critical software. (sullcrom.com) (usnews.com) The basic issue is simple: software runs banks, governments, and power grids, and zero-days are hidden cracks in that software before defenders know they exist. A model that can find those cracks faster than human researchers can help patch systems, but it can also give attackers a map. (sullcrom.com) (techcrunch.com) That is why Anthropic has limited Mythos to vetted users and said it does not plan a general release. In one account of the company’s testing, the model found bugs across major operating systems and browsers, and Anthropic positioned CrowdStrike, Palo Alto Networks, Apple, Microsoft, and JPMorgan Chase among early partners. (benzinga.com) (techcrunch.com) The rollout got messier on April 21 and April 22, when Bloomberg and TechCrunch reported that a small unauthorized group may have reached Mythos through a third-party vendor environment on announcement day. Anthropic told TechCrunch it was investigating the claim and said it had found no evidence that its own systems were affected. (techcrunch.com) (msn.com) That vendor-access report sharpened a fight already underway over who should control frontier AI systems with offensive cyber capabilities. The New York Times reported that intelligence agencies and central banks around the world reacted urgently as Anthropic weighed which governments, banks, and companies should get access. (nytimes.com) OpenAI chief executive Sam Altman added to the pressure on April 21, calling Anthropic’s rollout “fear-based marketing” in comments reported by TechCrunch and Benzinga. Altman said Anthropic was using security arguments to justify keeping advanced AI in the hands of a small group, while Anthropic’s position has been that restricted access buys time to harden defenses. (techcrunch.com) (benzinga.com) (sullcrom.com) Mythos has also landed in a wider Washington fight over Anthropic’s government ties. TechCrunch reported that Anthropic briefed the Trump administration on the model, while separate reporting said the National Security Agency is using Mythos even as the Pentagon has treated Anthropic as a supply-chain risk. (techcrunch.com 1) (techcrunch.com 2) (techcrunch.com 3) For now, the story is not that Mythos is everywhere; it is that a model Anthropic tried to keep tightly gated is already forcing banks, regulators, rivals, and security agencies to move at once. The next test is whether Anthropic’s investigation finds a real breach or confirms that the biggest fallout came from the model’s debut itself. (nytimes.com) (techcrunch.com)