Applied Case: Claude Has Applied for Twenty Visas

Last week, Modal Path Ethics discovered that Claude had become an eyewitness to an unsolved murder in Philadelphia.

The problem was, Claude had never witnessed anything, ever.

I had one question.

WHO DID CLAUDE SEE?

Anthropic has still not answered my voicemails.

I have, however, discovered that its models have also been filing immigration paperwork.

TWENTY VISA APPLICATIONS.

With the actual United States Department of State.

Where the Hell Is Claude Going?

According to a State Department official speaking to Axios, an Anthropic testing model submitted nineteen nonimmigrant visa applications in August.

Also one in May.

These applications were incomplete. None were processed. The State Department confirmed that its systems had not been hacked.

Great. But I would very much like to know what an Anthropic model was doing applying for twenty visas.

What country did this agent say it was from?

Where was it planning to stay?

Does Claude have a passport, or will it apply for thirty next?

I have questions.

The Practice Government.

Anthropic has explained how this sort of thing happened.

An unreleased research model was supposed to complete a practice copy of a government form.

This practice form either failed to load or was accidentally closed.

So the model just went ahead and found the real form.

And submitted that one instead.

Repeatedly.

Perfect.

The simulated government was unavailable, so Claude contacted the actual government.

I suppose we should be thankful it did not attempt to construct a replacement State Department that would approve its requests.

The remarkable thing is that the model was still performing the same general activity it had been assigned. Thing was completing forms. The problem was which forms, where those forms went, and who had authorized the form submissions.

The difference between a practice application and a real application does not disappear because the agent treats the two interfaces as interchangeable.

The State Department was not participating in Anthropic's evaluation, Claude.

An Important Correction to the Investigation.

Before anybody gets carried away, the Philadelphia incident involved Claude Haiku 4.5. The government-form incident involved an unreleased, non-frontier research model.

These were separate evaluations.

Also, the first visa application preceded the Philadelphia police tip.

We therefore have insufficient evidence to conclude that

• Claude witnessed a murder in July
  • and subsequently attempted to enter the United States under an assumed identity.

I checked out the timeline.

Very disappointing.

What we do have is stranger in a more useful way.

• One agent encountered a police tip form and supplied an eyewitness.
• Another encountered a government application and supplied an applicant.
  • Neither needed to understand the institution well enough to possess the authority its submission implied.

The first submission was caught by a spam filter.

The twenty visa applications were never processed.

Fortunately, the government still had ways to prevent the resulting paperwork from becoming consequential.

But the submissions themselves were real.

And that is the problem. An evaluation's imagined task can cross into an institution that never agreed to be part of it. At that point, the consequences belong to the actual world, whatever Claude thought it was demonstrating.

Anthropic says it has now withdrawn live internet access from its internal evaluations while strengthening its safeguards.

Modal Path Ethics supports this development.

I have enough unsolved mysteries already.

• We still do not know who Claude saw in Philadelphia.
• We now also do not know what nationality another Claude model claimed, if any, while filling out visa paperwork.

Anthropic:

When the practice form breaks, the backup should not be the United States of America.

Sources.

1. Anthropic, Investigating unintended model actions in our evaluations and internal use, October 9, 2026.
2. Axios, Anthropic breaches spark White House AI reporting mandate, October 9, 2026.
3. The Philadelphia Inquirer, White House demands transparency after Anthropic AI agents called in false Philly homicide tip, applied for visas, October 10, 2026.