Anthropic Reveals Claude AI Misuse in 150-Page Report

Anthropic published a threat report Wednesday running more than 150 pages, cataloguing eight months of attempts to misuse Claude. The document reads less like a corporate transparency post and more like a case file. Rocket guidance requests.

Bioweapons queries. Espionage attempts routed through the model. And a wrinkle that will irritate competitors more than regulators: Chinese labs training their own systems on Claude's outputs, then reselling access to customers as if it were their own product.

The company frames the report as evidence its safety filters work, since every case listed was one Anthropic caught and shut down. But the sheer range of misuse attempts tells its own story about where Claude sits in the world right now: valuable enough to steal, powerful enough to weaponize, and widely used enough that state-linked actors are trying to launder its capabilities into their own offerings.

None of this is speculative. It's Anthropic's own accounting, released on its own timeline, of what happens when a frontier model reaches global scale. The specifics are the point.