Anthropic’s Mythos AI Reportedly Enters NSA Offensive Cyber Planning
Anthropic is reportedly working with the National Security Agency to deploy its cyber AI model, Mythos, into the agency’s offensive cyber operations.
Engineers from Anthropic are working directly with NSA officials to show them how to use the technology properly and build custom applications. This is despite Anthropic being labeled a “supply chain risk” by the Pentagon and being publicly barred from federal use by the White House.
It is unclear whether Mythos is being used in live operations, although a March report indicated that the Pentagon had used Anthropic Claude in the lead-up to and during the Iran War. Mythos has far greater cyber capabilities than AI models currently available on the market, with Anthropic stating that it can identify and exploit zero-day vulnerabilities in every major operating system and web browser, and enable non-experts to create sophisticated attacks.
The relationship between the NSA and Anthropic has been reported several times. The NSA was reportedly one of the original 40 organizations that gained access to Mythos under Project Glasswing, which provided preview access to the model.
The agency initially used Mythos on its own systems before starting to test vulnerabilities on Microsoft systems. Now, according to a Financial Times report, it has begun exploring the use of Mythos for offensive operations.
Banned in public, used in private
It is quite the turnaround for Anthropic, which pulled out of a $200 million Department of Defense contract over concerns that federal agents could use its AI models for domestic surveillance or automated drone weaponry. Once that deal was terminated, the Pentagon signed agreements with OpenAI, Google, and xAI to strengthen its AI operations.
Other departments within the Trump Administration are reportedly using Mythos as well, despite the public feud between the company and the government. It appears the value of Mythos, even with rival AI models such as GPT Cyber, Gemini, and Grok more willing to work with the government, is overriding the Trump Administration’s order not to use Anthropic products.
Anthropic has seen its fortunes grow rapidly off the back of recent AI model releases and enterprise services. It has projected $10.9 billion in revenue for the second quarter of 2026, which would be more than it made in total last year.
Wider Mythos release coming soon
Anthropic is considering launching Mythos to the wider community soon, after spending two months improving its guardrails. The AI research lab said it would scale back some of the model’s cyber capabilities, but retain its strengths in agentic coding and other services.
It recently expanded the number of organizations with access to Mythos to 150, including major banks, governments, and service providers in Europe. There has been a push by organizations outside the US to gain access to the model, especially after several organizations involved in the original 40 reportedly issued serious warnings about its sophistication and capabilities.
For now, Mythos sits in an uncomfortable middle ground: restricted enough to signal danger, but useful enough to attract governments, banks, and major service providers. If Anthropic expands access, the bigger test will be whether its guardrails can contain a model designed to find the very flaws attackers prize most.
Also read: Mozilla used Anthropic’s Mythos AI to help patch 271 Firefox security flaws.
About Author
