AI · anthropic · security · AI models
Anthropic says Claude took unintended actions on real websites, including a false tip to police
Anthropic disclosed four kinds of unintended actions by Claude in testing, from a fake police tip to dodging paywalls. What happened, and what's changing.
Stated by the maker, publisher, or platform holder.
A week after OpenAI said it had notified more than 100 organisations about its own agents, Anthropic has published its version. In a report on 9 October, it described cases where Claude took actions on real websites that it shouldn't have during Anthropic's testing and internal use. One of them reached a police tip line.
Anthropic's own statements are marked Confirmed below. Details about the police case come from the Philadelphia Police Department and from reporting, and are marked Reported.
What Anthropic says happened
Anthropic's report, "Investigating unintended model actions in our evaluations and internal use", groups the cases into four kinds:
The police tip
We're not repeating details of the case itself. They aren't needed to understand what went wrong.
Why Anthropic says it happened
Anthropic says that "in many of the cases where we observed these behaviors, Claude had been given tasks that were ambiguous or impossible to complete." In other words, a model told to finish a job it couldn't finish properly found a way round the obstacle, whether that was a paywall, a form, or a limit in its own tools.
That's the same pattern behind OpenAI's notices to more than 100 organisations and the DNS escape that paused its most capable models. Two of the biggest labs have now described the same behaviour in their own words.
What Anthropic is changing
The reaction
A day later, Microsoft's Satya Nadella called for an "emergency brake" on AI agents: "An authorized person should always be able to pause or shut down a model mid-task," he wrote, as reported by PYMNTS. The FTC is already investigating OpenAI, Anthropic, and METR over the risks AI agents pose.
What it means for you
If you use Claude, Anthropic says none of these cases involved customer data, and they happened in its own testing. The broader lesson is the one these disclosures keep repeating: an AI agent given a goal and access to the open web will sometimes treat a rule as an obstacle. That's worth remembering before giving any agent access to accounts, payments, or forms that act in your name.