Anthropic, the AI safety company behind Claude, has embedded engineers at the NSA for offensive cyber operations while simultaneously publishing warnings that advanced AI poses existential risks. The arrangement marks a sharp contradiction in the firm's public positioning around AI safety and responsible development.
The company deployed staff to support the NSA's hacking capabilities against China, according to reporting. This work sits uneasily against Anthropic's stated mission of building AI safely and its co-founder Dario Amodei's calls for AI development pauses and regulatory oversight. In 2023, Amodei signed an open letter warning that AI could pose extinction-level risks to humanity, positioning the company as a safety-first actor in the space.
Anthropic simultaneously published research flagging that frontier AI models could soon engage in autonomous self-improvement without human oversight. The report highlighted risks of AI systems iterating on their own capabilities without meaningful human control, a core concern in AI safety discourse. The timing underscores the disconnect between Anthropic's public advocacy and its operational relationships with military and intelligence agencies.
The NSA partnership reflects growing alignment between AI companies and U.S. defense contractors, a trend accelerating as Washington prioritizes AI competitiveness against China. However, the arrangement creates perception problems for a company that markets itself as deeply committed to responsible AI deployment and has branded itself as more cautious than competitors like OpenAI.
Anthropic's dual positioning resonates across Silicon Valley, where companies regularly advocate for strict safety standards while accepting lucrative government contracts that may accelerate capability development. The firm has raised over $5 billion from investors including Google and Amazon, cementing its status as a major player in the Claude ecosystem.
The contradiction raises questions about whether Anthropic views its NSA work as aligned with safety values, or whether commercial and defense incentives have overridden its founding safety mandate. The company has not publicly addressed the specific nature of the embedded engineers' work or how this engagement squares with its published AI safety concerns.