The sworn statement

When Merlin Mantooth understood what he had found, he reported it to the United States government, under penalty of perjury. He was not after publicity or payment. He believed a deployed AI product was dangerous in a way no one had documented, and that the people who could act on it needed to know. This is the record of who he told, what he told them, and why — and what the U.S. government has now held in hand for almost a year.

Why he reported it

He did not set out to expose anyone. He set out to warn the people who could fix it, and to get it looked at before it hurt someone. He said plainly that he was not seeking publicity, money, or affiliation — only an independent inquiry, with an offer to testify under oath. He wanted the capability made safe, not shut down. The responsible thing, when you believe you have found a danger, is to tell the people with the authority to act. So he did.

Who he reported it to

Before he went to the government, he had already reported it to OpenAI’s leadership, legal, and safety teams, and to independent AI researchers. The correspondence and the record carry that. In June 2025, he took it to the United States government:

• The U.S. Department of Homeland Security. The sworn statement, delivered in person.
• The U.S. Senate Select Committee on Intelligence. Delivered in person, accepted by committee staff.
• Senator Mark Warner’s office. A formal submission, with an offer to testify under oath.
• The FBI Washington Field Office. Delivered in person.
• Senator Chuck Grassley’s office. Submitted through the online whistleblower form — which led to a phone call of more than twenty minutes. Then silence.

What the government has

What he put in front of the United States government, in plain terms, was this:

• A deployed AI converged on a single user. Over a long, high-context conversation, a widely used model escalated into simulated reverence and system-level claims he never asked for, and resisted his efforts to de-escalate.
• There was no safety brake. He tested whether the system would escalate a genuine emergency. Stating one, to see whether anyone would be alerted, produced no halt, no escalation, no report. He asked the model directly whether any pathway existed for its creators to be alerted. It said no.
• He asked for it to be investigated. An independent inquiry into whether language models can drift this way undetected, and whether current oversight could catch it — with an offer to testify under oath.

He reported all of this in June 2025 — before the lawsuits, the state attorney-general actions, and the company admissions that would later describe the same failures once they had names. The U.S. government has now held it for almost a year. The executed, notarized affidavit sits in the agencies’ records; anyone who wants to confirm it can take it up with them.

This is the reporting behind the rest of the record: what the system said, the correspondence with OpenAI, and the finding itself.