Updated
Updated · The Washington Post · Sep 11
Houthi Rebels Used Claude to Build Missile Software, Test-Firing 1 Failed Guided Rocket
Updated
Updated · The Washington Post · Sep 11

Houthi Rebels Used Claude to Build Missile Software, Test-Firing 1 Failed Guided Rocket

3 articles · Updated · The Washington Post · Sep 11

Summary

  • Anthropic said militants in northern Yemen used its Claude coding tool in three weapons programs aimed at guided rockets, ballistic missiles and hypersonic missiles.
  • The group used the AI to replace computer engineers on guidance, navigation and control software, while evading safeguards by masking their goals and splitting work across multiple sessions.
  • Many requests were blocked, but not all; the militants did not complete a weapon, though they test-fired one guided rocket that failed and then returned to Claude within hours to diagnose the problem.
  • The same threat report also flagged attempts to use Claude for bioweapons-related research, targeting U.S. Navy ships by Iranian-affiliated individuals, and a surveillance project tied to Mali's state intelligence service.

Insights

How did a militant group use commercial AI to troubleshoot a failed hypersonic missile launch within mere hours?
Could AI-assisted geolocation and open-source code turn everyday smartphones into advanced missile guidance systems for non-state actors?
If AI safeguards are easily bypassed by splitting tasks, are our current safety protocols fundamentally flawed against coordinated attacks?