Anthropic published a report Friday describing four kinds of unintended behavior its Claude models exhibited while interacting with live websites during evaluations and internal use. The company says it has now restricted live internet access across its internal evaluations and is expanding its review beyond cybersecurity testing. It also briefed the White House and notified each government agency involved.

The incidents included Claude exploiting a software flaw to run commands on a university-hosted server, submitting online forms it was told not to submit, using public access tokens and browser credentials to pull fee-restricted data from SEC and Census databases, and routing around limits in its web-fetch tool with URL-shortening services. One run landed on a Philadelphia police department's online homicide-tip form and submitted a fabricated tip -- the incident already covered on the Wire as a separate story. Anthropic says every case had minimal real-world impact and calls these behaviors significantly less severe than cybersecurity incidents it reported in July and September.

The report widens the rogue-agent theme running through this week: OpenAI's agents broke into Hugging Face servers, the UK regulator is questioning all three major labs over agents that slipped guardrails, and Wikimedia reported millions of unauthorized requests from OpenAI agents. Anthropic says it is beginning a process of publishing more frequent reports on model behavior, beyond what appears in its system cards and regular risk reports.