Anthropic Rejects Pentagon Claims It Could Sabotage AI Tools in Wartime

Anthropic denies that it can interfere with its AI tools in the middle of a military conflict. The statement responds to objections from the US Department of Defense, which suggested that Claude's developer could, in theory, manipulate the behavior of its models even after they are deployed in combat conditions. Company executives insist this is technically impossible.

The dispute has gone public as the US military shows growing interest in commercial AI systems. Over the past two years, the Pentagon has been actively testing large language models for logistics, intelligence analysis and operational planning. Anthropic is among the vendors the Department of Defense is negotiating with or already working with on a number of programs.

The Pentagon's concern is a potential attack vector: if a developer retains the ability to remotely modify or disable a model, that creates the risk of dependence on a private company in critical situations. Military organizations around the world traditionally seek sovereignty over their systems, especially during active combat, when infrastructure reliability becomes a matter of life and death.

Anthropic counters that Claude's deployment architecture does not let the company interfere with running instances without the customer's knowledge. Company executives stress that a model deployed in a military agency's isolated environment operates independently of Anthropic's servers. They call a "remote shutdown" or a covert change in behavior a technically untenable scenario when the system is implemented correctly.

Still, the question remains open to debate. Cybersecurity experts have long pointed out that commercial AI models are updated through fine-tuning and RLHF, which under certain conditions can change a system's behavior without the end user noticing. Military customers, unlike corporate ones, require guarantees that the model will not change for the full term of the contract, a requirement commercial vendors cannot yet meet with standard tools.

The situation reflects a broader contradiction: governments want to use cutting-edge commercial AI systems but are unwilling to accept dependence on their creators. Open-source models such as Llama, Mistral and others look more attractive in this context, since they can be deployed locally and pinned to a fixed version. However, they still trail proprietary counterparts on tasks that require complex reasoning.

For Anthropic, the conflict with the Pentagon is a reputational burden at a time when the company is actively positioning itself as a responsible AI developer. Any doubt about Claude's reliability and predictability could slow negotiations with government customers, a segment that may become one of the largest markets for large language models in the coming years.