IC Technology & Modernization — 2026-03-27

Project Maven Runs on 160+ Intelligence Feeds and Anthropic Claude in First AI-Fueled War, Pentagon Investigating School Strike

NPR Fresh Air revealed that the Maven Smart System integrates 160+ intelligence feeds and uses Anthropic's Claude LLM for target sign-off processes, course-of-action planning, and weapons-target pairing, everything short of the final human decision. CENTCOM has struck 9,000+ targets using Maven since operations began. The Pentagon is investigating whether Maven contributed to the Day 1 Minab school strike that killed 165+, which used coordinates more than a decade old. NGA officials acknowledged AI sycophancy and escalation risks, claiming safeguards are built "under the hood." Meanwhile, Anthropic's blacklisting for refusing autonomous weapons use has created an unresolved gap: the company that built the classified cloud AI is now banned, while OpenAI, which had no classified cloud presence, stepped in with potentially different safeguards.

Analysis
The Minab school strike, 165+ dead from decade-old coordinates, is the first potential AI-adjacent targeting failure at this scale. Maven's Claude integration creates a paradox: Anthropic built the classified cloud AI now powering US targeting, but Anthropic was subsequently blacklisted for refusing autonomous weapons use. NGA officials acknowledge sycophancy and escalation bias risks but claim safeguards are built in. With 9,000+ strikes executed through Maven, the system's failure modes are no longer hypothetical.

1 sources

America's first AI-fueled war is unfolding right now in Iran. This is how we got here - NPR