US security agency using Anthropic's AI despite blacklist
What's the story
The US National Security Agency (NSA) is using Anthropic's advanced Mythos Preview model, according to Axios. This comes even as the Department of Defense (DoD) considers the company a "supply chain risk." The development highlights a growing divide within the US government over the role of advanced artificial intelligence (AI) tools in national security operations.
Tool details
Mythos Preview's capabilities
Mythos Preview is Anthropic's most advanced model yet, boasting strong cybersecurity capabilities. Despite concerns raised by defense officials over the company, the NSA seem to be prioritizing operational needs and moving ahead with its adoption. The DoD had previously moved to restrict the use of Anthropic's tools, reportedly directing vendors in February to limit engagement with the company over a "supply chain risk."
Dispute details
Contradictory stance within US government
Despite the DoD's restrictions, sources told Axios that the NSA has continued to use Mythos, with some usage extending across other parts of the department. This apparent contradiction highlights a broader tension within the US government, with military officials questioning Anthropic's reliability in court while some elements within its ecosystem rely on its technology.
Negotiation details
Tension over contract renegotiations and defense requirements
The disagreement between the DoD and Anthropic stems from contract renegotiations earlier this year. Defense officials had pushed for broader access to Anthropic's models for "all lawful purposes," but the company resisted certain applications, particularly mass domestic surveillance and autonomous weapons. Some Pentagon officials view this stance as a liability, raising doubts about whether Anthropic can fully support defense requirements when needed.
Access control
Limited access to Mythos and potential international reach
Anthropic has tightly controlled access to Mythos, limiting availability to some 40 organizations over concerns of its potential misuse in offensive cyber operations. While 12 of these partners have been publicly identified, sources suggest the NSA is among those granted access. The model's reach may not be limited to the US as counterparts in the UK have indicated they can access similar capabilities through national AI security initiatives.
Discussion outcomes
Productive talks between US officials and Dario Amodei
Recently, Anthropic CEO Dario Amodei met senior US officials, including White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. The talks were centered around Mythos deployment and broader AI security considerations. Sources say the discussions were productive and could shape how agencies beyond the Pentagon engage with the technology in future.