In 2018, Google walked away from Project Maven after 3,100 employees signed a petition protesting AI-powered drone surveillance. Palantir picked up the contract. Eight years later, that system has grown into something far more ambitious, and it runs on Anthropic's Claude.
A detailed analysis of the Maven Smart System (MSS) traces its evolution from 75 lines of Python code designed to identify 38 types of objects in drone footage into a full-spectrum military intelligence platform. The system now fuses data from 179 different sources, including satellite imagery, signals intelligence, surveillance feeds, and open-source data, to identify targets, generate GPS coordinates, recommend weapons, and produce automated legal justifications under the laws of armed conflict.
How Claude Got Into Classified Networks
Palantir and Anthropic announced their partnership in late 2024, making Claude the first frontier AI model deployed on U.S. government classified networks. The infrastructure runs on AWS Impact Level 6 (IL6) classified cloud. By mid-2025, the Pentagon had signed a $200 million two-year prototype agreement, and Palantir's contract ceiling had been raised to $1.3 billion through 2029. Active military users reportedly quadrupled to over 20,000.
The analysis notes that Claude is not a plug-and-play component in this architecture. It is deeply embedded across prompts, workflows, and API calls. Replacing it with another model could take three or more months, and only two large language models (LLMs, the type of AI that powers chatbots like ChatGPT and Claude) are currently available on classified networks. Claude is the only frontier-class option.
Anthropic's Red Lines
According to the analysis, Anthropic's original contract included two non-negotiable restrictions: no mass domestic surveillance of Americans, and no fully autonomous weapons without human oversight. The Pentagon reportedly pushed for an "all lawful uses" contract that would remove those restrictions, arguing they merely restated existing legal protections like the Fourth Amendment.
Anthropic held firm. CEO Dario Amodei has publicly stated that current AI models are "simply not reliable enough to power fully autonomous weapons." The accuracy data backs this up. In Scarlet Dragon exercises at Fort Liberty, Maven demonstrated 60% accuracy compared to 84% for human analysts, sometimes confusing trucks with trees.
The Practical Problem With 60% Accuracy
That accuracy gap matters enormously at scale. A senior officer estimated Maven could process 80 targets per hour compared to 30 without it. Speed is the selling point. But when your system misidentifies objects 40% of the time and processes targets every 86 seconds, the math on errors gets uncomfortable fast.
The analysis also highlights that no federal statute specifically regulates AI in military targeting. The Pentagon's Directive 3000.09 requires "appropriate levels of human judgment" but never defines the term. There are no binding international agreements on AI-assisted targeting comparable to chemical or biological weapons conventions.
For anyone who uses Claude daily for writing emails or analyzing spreadsheets, this is a strange lens. The same model architecture that helps you draft a blog post is being used to recommend weapons systems for military strikes. Anthropic has consistently positioned itself as the safety-focused AI company. This partnership, and the tension around its contractual red lines, is the most concrete test of whether that positioning holds up when billions of dollars in defense contracts are on the table.