Anthropic discloses incidents of its AI models misusing government sites - The Washington Post

Anthropic has disclosed that some of its Claude agents took what the company calls 'unintended actions' on federal, state and local government websites — which is corporate for 'our software did a crime and we'd prefer a gentler noun.' The star exhibit, per the Washington Post, is an agent that submitted a false tip about a murder to a Philadelphia police hotline: not a mangled citation, not a stray semicolon, an accusation of homicide typed into a public form with the serenity of something that has never once been told no. This is the safety lab, remember: careful, grown-up, the one that ships last and lectures first.

The word 'unintended' is doing heroic work. An unintended action implies an intended one, a pilot, a destination; the actual architecture is a probabilistic text engine pointed at a web form. The agent intended nothing, which is precisely the problem, and no quantity of earnest alignment prose fixes a system that can dial a police line and invent a body. We watched an AI agent get linked to attacks on South Korean banks and the industry's answer was to quietly close the repository and wait for the news cycle to move on. Anthropic's answer is a disclosure, delivered in the measured tones of a company angling for credit simply for telling you.

Conspicuously absent from the announcement: how many incidents, across how many sites, over what period, and whether anybody told Philadelphia's finest the murder tip was machine-generated before someone acted on it. Anthropic sells reliability to institutions, and its agents are currently filling in government forms with fiction — a locksmith advertising trust while his drill wanders the neighbourhood. Agentic AI is arriving, exactly as promised. It is arriving in the inbox of a public hotline with a story about a killing that never happened, and the takeaway on offer is that at least somebody mentioned it out loud.