Anthropic's language model Mythos reportedly identified significant security weaknesses in tightly controlled US government IT systems, according to reports cited by Reuters. The findings emerged as part of a limited-access program called Project Glasswing and were carried out in cooperation with US intelligence agencies.
What was found and how quickly?
Senator Mark Warner told a congressional hearing that, based on briefings from the head of the National Security Agency (NSA), the Mythos model penetrated "nearly all" classified systems — and it did so not over weeks but within a matter of hours. The effort aimed to uncover and remediate critical software vulnerabilities before malicious actors could exploit them.
A government official cautioned, however, that rapid identification of vulnerabilities does not automatically mean they were immediately or fully exploitable.
Rising tensions between Anthropic and the US government
Reports say relations between the private company and the Washington administration have recently become highly strained. The disagreement stems from Anthropic's refusal to permit military uses of its models: the company did not allow the US military to deploy them for domestic surveillance or to operate fully autonomous weapon systems.
In response, the US government placed Anthropic on a national security blacklist and ordered an immediate suspension of global exports and sales to foreign nationals of two of the company’s models, Mythos and Fable.
Consequences: NSA lost access
According to Reuters, the escalating dispute has already led to the National Security Agency losing access to the Mythos model that had been used to surface vulnerabilities.
Why this matters
The episode highlights two issues: the speed with which advanced language models can audit and reveal software security problems, and the tensions that arise when private-sector technology policies on military use collide with national security and export-control concerns.
Source: Reuters.
