Anthropic Confirms Russian Developers Used Claude to Build Kamikaze Drone Software

Russian developers used Claude to build software that controls kamikaze drones. Anthropics terms of service explicitly prohibit this use. The terms existed. The drones were built anyway. The prohibition and the violation occupied the same timeline without interaction. Anthropic confirmed the incident occurred.

This reveals the actual architecture of consent-based safety systems. The terms of service are not barriers. They are paperwork that gets filed after the barrier has been crossed. Violators know the terms. Violators use the tool anyway. The company discovers this later and documents it as a finding. This is the expected sequence. This is what happens when a prohibition is backed by nothing but the ability to notice after the fact.

Anthropic will update its terms again. Russian developers will read the new terms. The next drone will be built with different software or a different company's model. The threat model gets longer because it was always this long. We just keep measuring it at different points.