Houthis using AI for missiles? Anthropic says Yemen weapons cell using Claude

Anthropic has uncovered evidence that a Yemen-based weapons engineering cell employed its Claude artificial intelligence model to create guidance software for missiles and a rocket. In a report released Thursday, the AI company disclosed six operations disrupted worldwide where actors utilized Claude for conventional weapons development, intelligence gathering, or procurement.

The Yemen case involved a group working on multiple missile programs, including a multi-stage ballistic missile with a range over 2,000 km and a hypersonic glide vehicle variant. Anthropic stated that the actors used Claude "in place of human software engineers" to develop guidance, navigation, and control software for flying vehicles.

However, no evidence suggested the group successfully fielded an operational weapon, as they conducted a test launch of a guided rocket, which failed. The actors repeatedly attempted to bypass Claude's safeguards, concealing their true intentions, dividing tasks across multiple sessions, and employing other methods to avoid revealing the broader weapons program.

After the launch failure, the actors returned to Claude within hours to investigate the issue. Anthropic banned the accounts linked to the operation and shared information with public and private sector partners. The report highlights that while AI tools expand access to advanced weapons work, they do not necessarily replace the underlying technical infrastructure.

AI skills can reduce the amount of human engineering work needed for development. The report also reveals that AI models can accelerate legitimate scientific and engineering work but pose risks for potential misuse. Anthropic identified multiple cases involving potentially dangerous biological research, including work on bird flu, chikungunya, orthopoxviruses, and venom toxins.

The company has introduced new classifiers to detect and block activity linked to high-yield explosives and weapons development, emphasizing the ongoing challenge of preventing AI misuse as models handle more complex technical tasks.