Anthropic has assigned engineers to the NSA to enhance cyber capabilities while advocating for a global pause in AI development amid rising concerns.
Anthropic Engages with NSA for Cyber Initiatives
Anthropic has deployed approximately six engineers to the National Security Agency (NSA) to leverage its advanced AI model, Mythos, for offensive cyber operations. This collaboration aims to enhance capabilities for potential actions against networks in nations such as China and Iran.
The engineers are tasked with tailoring the AI model to meet specific operational requirements, although it remains uncertain if they are currently involved in active missions. Mythos, which Anthropic has withheld from public release due to fears of misuse, is only available to approved partners, including major tech companies like Microsoft, Apple, and Amazon, through an initiative known as Project Glasswing.
Legal Challenges and AI Moratorium Advocacy
In a separate development, Anthropic is taking legal action against the Pentagon after being labeled a supply-chain risk by Defense Secretary Pete Hegseth, a designation typically reserved for foreign entities. This followed the collapse of a $200 million contract, which was partly due to Anthropic's refusal to allow the Department of Defense to use its AI, Claude, for autonomous weaponry or surveillance, although the NSA's usage of Mythos is exempted from this restriction.
A California judge has temporarily blocked the blacklisting of Anthropic, citing potential First Amendment violations. However, a D.C. appeals court has denied Anthropic's request to stop the blacklisting while the case is still pending, allowing the NSA to continue using Mythos.
AI Growth and Global Pause Proposal
On the same day as the NSA news, Anthropic released a report titled "When AI Builds Itself," which explores the advancements of its AI model, Claude, in automating software development. The report reveals that Claude is now responsible for generating over 80% of the code in Anthropic's production codebase, a marked increase since the introduction of Claude Code in early 2025.
Marina Favaro, a lead author of the report, along with co-founder Jack Clark, cautions that the rapid progress in AI could lead to recursive self-improvement, where AI systems independently create and refine their successors, significantly reducing human oversight.
The report highlights a pivotal moment in AI development, where systems begin to autonomously determine the value of experiments rather than merely executing tasks as directed by humans. This shift poses significant risks, as misalignments could propagate through self-improving iterations.
In response to these concerns, Anthropic advocates for a verifiable global pause in AI advancements, suggesting that simultaneous halts by leading labs, monitored for compliance, could mitigate potential dangers. The company acknowledges that a unilateral pause would be ineffective, as it would merely allow others to gain an advantage in the field.
Continued AI Development Amidst Concerns
This scenario reflects a broader tension in the AI landscape, where the same organizations warning about the risks of AI are also at the forefront of its development. Despite calls for caution, including an open letter from over a hundred AI experts in 2023 urging a collective effort to address the existential risks of AI, progress has continued unabated.
The Pentagon has set a deadline for August to withdraw Claude from its systems, coinciding with Anthropic's anticipated IPO, which is expected to bring greater scrutiny to its financial practices.
