Anthropic has sent engineers to the National Security Agency to work on offensive cyber operations, even as the company released a report cautioning that artificial intelligence could soon develop itself without human oversight. The dual moves place the AI safety startup in an unusual position — contributing to government attack capabilities while warning about the very technology it helps advance.
The NSA Placement
Engineers from Anthropic are now embedded at the NSA, according to the company. The role focuses on offensive cyber operations, meaning the engineers are helping the spy agency develop or deploy tools to penetrate adversary networks. Anthropic didn’t specify how many employees were involved or how long the assignment would last.
The deployment marks a rare, direct collaboration between an AI safety-focused firm and a U.S. intelligence agency. Most companies in the sector have emphasized defensive applications or voluntary safety pledges. Anthropic’s decision to embed staff inside the NSA for offensive work is a departure from that pattern.
Self-Building AI Warning
Separately, Anthropic published a report arguing that AI systems may soon be able to improve and reproduce themselves without any human input. The paper warns that such self-building capability could emerge faster than expected, potentially making it difficult for researchers to maintain control. The company urged the AI community to prepare safeguards now, rather than after the fact.
The report does not specify a timeline for when AI might achieve this milestone. But the warning is stark: if machines can write their own code and train new models, the pace of advancement could accelerate beyond human oversight. Anthropic’s own engineers are simultaneously working on offensive cyber tools, raising questions about where the line between development and deployment lies.
No Immediate Comment on Balance
Anthropic has not explained how the two efforts fit together. The NSA work pushes the company deeper into government offense, while the report calls for restraint. Critics might ask whether building attack tools for the NSA undermines the very safety research Anthropic is known for. The company has not addressed that tension publicly.
The NSA declined to discuss the details of its collaboration with Anthropic. The agency often keeps its partnerships confidential, but the disclosure that engineers are embedded signals a level of trust unusual for a relatively young AI company.
What comes next is unclear. The warning about self-building AI points to a future where human involvement shrinks. For now, Anthropic’s own humans are busy inside the NSA — building the capabilities that might one day need to be contained.
