Anthropic unexpectedly withdrew access to its Claude Fable 5 and Claude Mythos 5 artificial intelligence models just days after they became available, citing serious security concerns raised by US authorities. The company announced the withdrawal only a few days after the models were released.
Why access was shut down
US authorities flagged significant security risks related to the Mythos model. Initial public details were limited, but subsequent reporting and a Senate hearing provided more information about the concerns.
Anthropic initially intended to restrict further use to US citizens, but because verifying citizenship proved impractical, the company opted for a complete shutdown. The move affected US allies as well and was implemented without prior warning.
Senate hearing and alleged capabilities of Mythos
According to a report citing the Senate Intelligence Committee hearing, the Mythos model reportedly penetrated nearly all encrypted systems managed by the NSA (National Security Agency) and US Cyber Command. Joshua Rudd, described in reports as the general leading those organizations, stated that Mythos accomplished this in a matter of hours.
The hearing also included comments that, although China is currently about a year behind in AI, that lead could be erased if Anthropic or other US companies reach a point where AI can autonomously create more advanced versions of itself.
Project Glasswing and partner testing
Rather than broad public release, Anthropic provided access to Mythos to roughly 200 selected partners through a project called Project Glasswing. Participants in testing included Amazon, Apple, Google, Microsoft, Nvidia, JPMorgan and the Linux Foundation.
Anthropic has said Mythos already identified several thousand vulnerabilities across different systems, including 27 vulnerabilities in OpenBSD, which is often regarded as one of the most secure operating systems. Those findings illustrate the model’s potential to uncover security flaws.
Impact: both risk and defensive benefit
Observers note the restrictions have a dual effect: they make life harder for cybercriminals and state-sponsored hackers, but they also limit researchers and defenders who would use such models to build stronger protections for their systems.
The Glasswing effort underscores how access to advanced AI becomes a strategic asset. Europe has taken note of this dynamic as well, increasing interest in technological sovereignty and reducing dependence on external actors for protection of critical systems.
Conclusion
Anthropic’s withdrawal demonstrates how quickly a technology can move from commercial release to a national security issue. Given the seriousness of the claims around Mythos, the company chose to revoke access entirely while further investigations, the results of the Senate hearing, and partner testing inform next steps.
