According to the Financial Times, Anthropic has embedded roughly half a dozen engineers at the National Security Agency (NSA) to deploy its Mythos AI model for offensive cyber operations, potentially targeting networks in China and Iran. Mythos is the same model Anthropic withheld from public release due to misuse risks, releasing it only to vetted partners like Microsoft, Apple, and Amazon through Project Glasswing. Meanwhile, Anthropic is suing the Pentagon after Defense Secretary Pete Hegseth designated it a supply-chain risk following a collapsed $200 million contract, partly due to Anthropic's refusal to allow fully autonomous weapons or domestic mass surveillance use of Claude. A California judge blocked the blacklisting, but an appeals court denied Anthropic's halt bid; the NSA continued using Mythos. On the same day, Anthropic's research institute published a report warning that Claude now writes over 80% of Anthropic's production code and is approaching recursive self-improvement—AI systems autonomously designing successors. The report cites an experiment where Claude agents recovered 97% of a safety gap versus 23% by humans, exercising research judgment autonomously. Anthropic proposes a verifiable global pause on frontier AI development, but unilateral slowdowns risk losing the lead. The IPO filing, potentially valuing Anthropic above $1 trillion, coincides with these developments.

Key facts

• Anthropic embedded ~6 engineers at NSA to deploy Mythos AI for offensive cyber ops.
• Mythos was withheld from public release; used only by vetted partners like Microsoft.
• Anthropic sues Pentagon over $200M contract collapse; court blocks blacklisting.
• Claude now writes >80% of Anthropic's production code; AI approaching self-improvement.
• Anthropic proposes global pause on frontier AI, but unilateral slowdown risks losing lead.

KeyAudit data perspective