(June 19): Some firms chosen early on by Anthropic PBC to test the Mythos AI model ahead of a wider release have preserved their access to a preview of the system, despite a US government order that led to the total shutdown of other versions.
Businesses including banks and technology firms are accessing Mythos Preview through Project Glasswing, a collective of roughly 200 organisations cleared by Anthropic to use the system to hunt for cyber vulnerabilities, people familiar with the situation said, asking not to be identified because the information isn’t public. Dragos Inc chief technology officer Jon Lavender said his company has access. Cisco Systems Inc confirmed it’s retained access too.
In testing, Mythos Preview proved so powerful — identifying thousands of vulnerabilities, including in every major operating system and web browser — that Anthropic limited its release in April to Project Glasswing members only. The startup eventually made Fable 5, a version of Mythos that the company blocked from carrying out cybersecurity tasks, available to the public instead.
But last week, Anthropic abruptly cut off access to Fable 5 to the public and halted the rollout of Mythos 5 (a less restricted version of Fable) to Project Glasswing members. It did so after receiving a letter from the US government ordering it to block both models from use by any foreign national without a licence, the most aggressive intervention yet involving frontier AI systems.
The shutdown sowed confusion across the Project Glasswing community with some member organisations saying initially that their access to Anthropic models was disrupted, people familiar with the situation said. It wasn’t immediately clear whether Mythos Preview would remain available to them given the US government didn’t explicitly address the Preview version in its order.
Anthropic didn’t directly address the preview version either in a blog post explaining its decision to disable Fable 5 and Mythos 5, saying only that “all other Anthropic models” wouldn’t be affected. The company didn’t respond to a request for comment.
See also: Intel rallies to record as Trump touts Apple chip deal
For now, the access represents an opportunity for industries to continue to build cyber defences against ever-advancing AI systems that are proving adept at surfacing previously undiscovered ways to hack them. Cisco, Amazon Web Services, JPMorgan Chase & Co were among the first members of Project Glasswing.
Amazon.com Inc, which made Mythos Preview available to other members through its Bedrock platform, declined to comment and referred instead to Anthropic.
Not even Project Glasswing was fully shielded from the effects of the government order, however. The European Union’s cybersecurity agency, known as Enisa, had been invited to join Glasswing ahead of the US government’s block. The agency was informed on Friday that it would no longer be given access, an agency spokesperson said.
See also: Amazon in talks to sell custom AI chips in bid to cut Nvidia’s dominance
It wasn’t immediately clear how Anthropic was determining access to individual Glasswing members. A European Commission spokesperson said Enisa met with Anthropic on Thursday.
Uploaded by Liza Shireen Koshy
