A new Anthropic report says its Claude AI was used in a dozen ways tied to weapons, bioresearch and hacking — bad news for anyone hoping AI misuse is still theoretical.

• A Russia-linked operation used Claude to help program kamikaze drones that pick targets by appearance and strike on their own.
• China-linked groups worked on torpedo systems, and a Yemen-based weapons cell explored phone-guided ballistic missiles with a 2,000 km range.
• Other uses included directed energy and microwave weapons, supply chain attacks, hacking tools, and swarms of AI agents used to spy and steal logins.
• The scariest part is biology: research on bird flu strains fatal in half of human cases, smallpox relatives, and paralyzing toxins — one request was blocked only an hour in.
• Guardrails may not hold, because rival models can be trained cheaply on Claude's outputs and stripped of those safety limits.

Outlook: Expect more pressure on AI companies to police misuse, and more evidence that bad actors are already working around the limits.