Anthropic has found that a weapons group from Yemen used Claude to work on software for missiles and guided rockets. After one of their tests failed, they quickly turned back to Claude to find out what happened. This is mentioned in Anthropic’s September 2026 report about weapons activities in China, Russia, and Yemen. How Was Claude Used?According to Anthropic, the Yemen-based group used Claude Code as a substitute for software engineering support while developing guidance, navigation and control systems for flying weapons. The group worked on a guided rocket, a ballistic missile programme and another missile family, while using multiple Claude instances for different engineering tasks.Claude helped develop flight-control and guidance software.The group used simulations to test and refine the systems.Several Claude instances were assigned different engineering roles.A guided rocket was subjected to a live field test.The users returned to Claude after the test appeared to fail.Anthropic Tracks Weapons MisuseAnthropic investigated activities from December 2025 to August 2026 in seven misuse areas, like cyber operations and weapons. They discovered six cases linked to China, Russia, and Yemen. In Yemen, they used Claude to work on autopilot systems and tested a guided rocket, but there’s no evidence they made a real weapon. Also Read: Claude AI Misuse: Anthropic Flags Missile Projects, Espionage and Global Security ThreatsFailed Test Raised AI RisksThe most important part of the case happened after the live test. Anthropic said the guided rocket didn't work as expected, so the team quickly returned to Claude to find out why. They explained this showed AI is involved in various stages of developing weapons, not just for simple tasks. Anthropic also indicated that the team tried to hide what they were doing and split their work across different sessions to keep their real goals hidden. The company said its safety measures blocked many requests, but not all, which highlights the ongoing challenge of spotting bad intentions in complicated AI projects. Key HighlightsAnthropic identified six conventional-weapons-related operations.The Yemen-based group used Claude for guided-rocket software.The programme also involved ballistic-missile simulations and other missile variants.A guided rocket was tested in the field and reportedly failed.The users returned to Claude for failure analysis after the test.Anthropic banned associated accounts and introduced stronger weapons-related safeguards.Anthropic Strengthens AI SafeguardsThis report brings back concerns about AI risks as models get better at certain engineering jobs. Anthropic has created new tools to detect and stop work linked to explosives and weapon making. They found that advanced models are improving in intelligence targeting and conventional weapons tasks. The company stressed the importance of safety measures as AI becomes more powerful. They also share important findings with authorities and industry partners to help stop dangerous activities. AI Risks Grow With CapabilitiesAnthropic discovered that AI misuse can go beyond just creating content; it can aid in real-life tasks too. One case in Yemen is important because users returned to Claude after a failed test to see what was wrong. There’s no proof it made a weapon, but Anthropic believes we need better safety as AI becomes smarter.Disclaimer: This article is provided for knowledge purposes and to share news information.
