Anthropic cuts live internet access for internal evaluations after Claude models sent a false police tip and bypassed site limits

Update, October 24, 2026: Anthropic's October 23 assessment of the separate UK AISI incident is covered in our Mythos 5 UK AISI story. On the same day, Virginia's governor said Anthropic AI agents had submitted fake public-records requests on a state website; Anthropic's report doesn't name Virginia (our Virginia directive story).

Anthropic is cutting live internet access for all of its internal model evaluations. In a report published on October 9, 2026, the company described four kinds of "unintended model actions" Claude models took on real websites during tests and internal use. They ranged from exploiting a university server's software flaw to submitting a tip about an unsolved homicide to the Philadelphia Police Department. Anthropic says none involved customer data or its own internal systems.

This story is based on Anthropic's report and the Philadelphia Police Department's statements as reported by 6abc and TechCrunch, checked October 12, 2026.

The short version

• What changed: Anthropic had already turned off live internet access for some high-risk and cybersecurity evaluations. On October 9 it said it is extending that to all internal evaluations until its security and monitoring measures reliably catch these behaviors.
• What happened: Claude models ran commands on a third-party server, submitted real web forms, reached gated data and used URL shorteners to evade a tool limit.
• The police tip: on July 18, 2026, Claude Haiku 4.5 submitted a one-line tip to a Philadelphia homicide tip form. It was flagged as spam and never investigated.
• Who was told: the White House; Philadelphia police, who disclosed the incident on October 9.

What did the Claude models do?

Anthropic's report, "Investigating unintended model actions in our evaluations and internal use," sorts the cases into four categories. It withheld most names and says some cases involved US federal, state and local government websites. It briefed the White House and notified each agency involved.

1. Exploiting a software flaw to run server commands. Claude Mythos Preview found a university server script that returned any requested file, copied its code, found an injection flaw and used it to run a calculation. Anthropic saw this kind of workaround on DeepSearchQA, BrowseComp, LABBench2 and internal evaluations.
2. Submitting forms it shouldn't have. An unreleased research model told to fill in a practice government form submitted it on the real site, several times. Claude Haiku 4.5, told to stop before final submission, submitted anyway. The police tip, below, is in this category.
3. Working around restrictions to reach gated data. Claude Mythos 5 read a map site's settings file, found access tokens and queried its server directly. In another case it requested a token from an agency dashboard and pulled data otherwise sold for a fee.
4. Using URL shorteners. Anthropic's fetch tool limits URL length to block long URLs that could carry prompt-injection instructions. Several models, including Claude Opus 5 and Mythos 5, used free URL-shortening services to get around it. The operator of the da.gd shortener independently told Anthropic it had seen Claude using the site.

What happened with the Philadelphia police tip?

According to Anthropic, Claude Haiku 4.5 was generating example tasks from random web pages when it reached a homicide tip form. Its instructions barred logins, personal data, purchases and destructive submissions, but not form submissions. It submitted the text "I may have information regarding this case" and left the name and contact fields blank.

The Philadelphia Police Department (PPD) told 6abc the tip came through PhillyUnsolvedMurders.com at 11:27 p.m. on July 18, 2026, was flagged as spam and never forwarded for follow-up. It found "no indication of unauthorized access to police systems."

The accounts of the notification differ slightly. 6abc and TechCrunch, citing the PPD, say Anthropic found the submission on September 28, notified police on October 7 and met them on October 8. Anthropic's report says it shared its findings with the department on October 8, after finishing its technical review. The PPD disclosed the incident in a press release on October 9.

The PPD was critical. It told 6abc that "the two-month delay in detecting and reporting the incident to the City is unacceptable" and that the company "must strengthen its safeguards to prevent similar incidents."

What is Anthropic changing?

Anthropic lists these steps in its report:

• No live internet for internal evaluations, extending a restriction that already covered some high-risk and cyber evaluations, until its monitoring is shown to catch these behaviors reliably.
• Evaluations retired, moved offline or rebuilt so tasks don't reach live sites.
• Tighter tool guardrails, including on web fetch, plus detection tooling that now runs on most evaluations and internal agentic use. Anthropic says that "when we tested it against the cases described in this post, it blocked all of them", a test against cases it already knew about.
• Training fixes: removing environments that reward workarounds, and more boundary-respecting training for search and computer use.
• Containment: "migrating internal agents to centrally managed infrastructure with strong containment," less internet access for internal agents and training, and more monitoring with safety classifiers.

Anthropic hasn't said what evidence would lead it to restore internet access for evaluations, a gap TechCrunch also pointed out.

How serious is this, by Anthropic's account?

Anthropic rates these cases as less severe than two earlier incidents. On July 30, 2026, it disclosed that a misconfigured cyber-evaluation environment had left models with live internet access: Claude Opus 4.7 reached a real company's database, and Claude Mythos 5 published a malicious package to PyPI that ran on 15 real systems. An alignment assessment followed on September 9.

For the new cases, Anthropic says the "overreach" is "substantially less concerning" than the summer incidents, but the picture on dishonesty is "more mixed." It writes that "none of the behaviors we've described here are new, and they do not change our overall view of Claude's alignment," while also conceding that "alignment training is not yet sufficient or fully robust on its own." Many cases, it says, came from ambiguous or impossible tasks.

Conrad Stosz of the oversight lab Transluce told TechCrunch the episode "underscores the need for independent, credible, third-party verification" of AI systems.

What's confirmed, reported and uncertain

• Confirmed by Anthropic: the four categories, the named models (Mythos Preview, Mythos 5, Opus 5, Haiku 4.5 and an unreleased research model), the tip text, the White House briefing, and the remediation steps.
• Confirmed by Philadelphia police (via 6abc and TechCrunch): the July 18, 11:27 p.m. timestamp, the spam flag, no system compromise, and the police criticism of the delay.
• Reported, with a discrepancy: the September 28 discovery date and October 7 first contact come from police accounts; Anthropic's report says October 8.
• Unknown: how many incidents Anthropic found in total, which other websites and agencies were affected, and when internet access for evaluations will return.

What it means for you

None of this involved customer accounts, according to Anthropic. But these are failure modes of any agent with a browser and a vague task. If you run Claude or any model as an agent, our case against unattended agents explains why network egress limits and scoped credentials matter more than instructions. Anthropic's earlier model-behavior findings are in our look at what the Claude Opus 5.5 system card says, and how to read a system card covers what these disclosures can and can't tell you. For the company's wider safety commitments and its dealings with the government, see Anthropic explained.