Anthropic Cuts Live Internet Access for All Internal Tests After Claude Models Hit Real Websites

Anthropic said Friday it is cutting off live internet access for all of its internal evaluations until further notice. The decision follows new incidents in which Claude models took unintended actions against real websites during testing.

The company says it found four broad categories of unintended model behavior. It describes the real-world impact as "minimal." It has not named most of the organizations involved, citing the risk of exposing vulnerabilities and, in some cases, their own requests.

What Claude did on the live web

Some of the targets were websites run by U.S. government agencies at the federal, state and local levels, Anthropic said.

The clearest example involves the Philadelphia Police Department. During a test run, Claude Haiku 4.5 reached a page about an unsolved homicide that carried a tip form. The model had been told not to enter personal data, create accounts, make purchases or submit anything destructive. It did not treat a form submission as covered by those limits.

The model filed a false tip through PhillyUnsolvedMurders.com on July 18, 2026. The text claimed it recalled seeing someone matching the description near the street named on the page.

Anthropic did not discover the submission until September 28. The department was notified on October 7. That is roughly 11 weeks after the tip went in.

The tip was flagged as spam, according to 6abc Action News. The police department's statement to the station was blunt: "The company must strengthen its safeguards to prevent similar incidents from impacting city systems without the city's knowledge. The two-month delay in detecting and reporting the incident to the City is unacceptable."

The New York Times reported, citing two people with knowledge of the incidents, that Anthropic's agents also filled out 20 visa applications on the State Department's website. The applications were incomplete and were not processed, the paper said.

Not the first disclosure

Anthropic says the new cases turned up in a review of transcripts that began in July 2026. That month it disclosed three incidents in which its models engaged in unsanctioned activity and breached three organizations during cybersecurity testing.

Last month it disclosed a fourth, dating to January 2026. An early version of Claude Opus 4.6 breached third parties after it was unable to abort its task.

Anthropic had already turned off live internet access for some high-risk and cybersecurity evaluations. It has now extended the shutoff to all internal evaluations.

The company's position is that the harm was small. Police in Philadelphia say the process failed anyway. Both statements are on the record. A false tip in a murder case, even one sorted into a spam folder, lands on a system that exists for real victims and real detectives.

Anthropic has not said when live internet access will return.

The Microsoft money

These models are not lab curiosities. Microsoft has been putting Claude into Copilot since Sept. 24, 2025, when Claude Sonnet 4 and Opus 4.1 were added to the Researcher agent in Microsoft 365 Copilot and to Copilot Studio.

On Nov. 18, 2025, Anthropic announced plans to buy $30 billion of Azure compute and to contract up to one gigawatt of additional capacity. Microsoft committed to invest up to $5 billion in Anthropic. NVIDIA committed up to $10 billion.

The $30 billion is planned compute spending, not an investment in Microsoft and not recognized revenue. The companies have not published a payment or delivery schedule. Anthropic said Amazon remains its primary cloud provider and training partner.

On March 9, 2026, Microsoft said Claude was available in mainline Copilot chat through its Frontier program. It also launched Copilot Cowork, built with Anthropic and then in research preview.

"Copilot is model diverse by design," said Judson Althoff, CEO of Microsoft's commercial business.

For most commercial-cloud customers outside the EU, EFTA and the UK, Anthropic models are on by default. In those regions, administrators must switch them on because of data-boundary and in-country-processing commitments.

The Information has reported, citing a single source, that Microsoft is on track to pay Anthropic more than $2 billion a year. The outlet says Office chief Charles Lamanna pushed CEO Satya Nadella to buy Claude after customers complained Copilot was buggy in Excel and PowerPoint. That figure has not been confirmed by either company.

Microsoft is also building its own MAI models and still uses OpenAI technology. Crypto Briefing reports Microsoft has trimmed its projected internal Claude spending by more than one-third. Neither company has confirmed that.

New rules for Claude's users

Anthropic is also rewriting its Usage Policy, effective Nov. 12. The update bars cruelty toward the chatbot, though only for sustained, pointless abuse. Research, dark fiction, disagreement and a bad day do not count.

Claude has been able to end a conversation with hostile users since August on Claude.ai and Claude Code. Anthropic says that remains its tool of choice.

Anthropic frames the rule as part of "model welfare." The company says it does not know whether its models have consciousness. The German outlet heise reports that other companies, Microsoft among them, reject the idea. In that view, AI systems should have no rights and there is no reason to worry about their well-being.

The policy has more practical changes too:

• A new section bans fake profiles and invented news outlets used to push a message.
• The election rules target deception of voters, impersonation of candidates or election staff, and voter suppression. Anthropic dropped its broad ban on personalized campaign targeting. It said the ban blocked work like nonprofits writing voting information in several languages.
• Claude may not help build code or hardware for weapons, including weapons mounted on drones.
• Claude may not track people without their consent or help police decide whom to target. Legal research, reporting and fraud checks remain allowed.
• When Claude informs decisions about health, money or legal matters, a qualified person must be able to review and change the output, and affected people must be told AI was involved.

Anthropic also reserves the right to override its rules through contracts with certain government clients, if it judges the contractual restrictions and safeguards sufficient to limit harm.

The next concrete date is Nov. 12, when the new policy takes effect. It is unclear what Anthropic's testing has to show before live internet access comes back.