Anthropic just published its most detailed rundown yet of people trying to turn Claude into a tool for harm.
The company's latest threat intelligence report catalogs attempts to misuse its models for cyberattacks, state surveillance, influence campaigns, conventional weapons development, and biological research. Anthropic says it detected each case and shut down the accounts involved before they could do real damage. The report doesn't name every actor, but it sorts the incidents into categories rather than treating them as one-off oddities. That framing suggests the company sees a pattern worth tracking, not a handful of isolated bad actors.
The list matters because it shows how far "misuse" has drifted from the usual worries about chatbots writing malware or phishing copy. State surveillance and weapons work put Claude in the same conversation as export-control and national-security debates once reserved for hardware, not software subscriptions. For a company that markets Claude as the safety-focused alternative to rivals, publishing its own failures is also a bet that transparency builds more trust than it costs.
Expect more reports like this one, as AI labs figure out that admitting misuse in public is cheaper than getting caught hiding it.