The short version
  1. Only one charge has been filed: attempting to intercept conversations with a device, with a maximum of five years. There's no espionage, trade-secret or foreign-agent charge, and no filing alleges a motive.
  2. AI chats are part of the evidence. The complaint cites queries to OpenAI's internal assistant and prompts from a chatbot account linked to the accused's spouse, who is not charged.
  3. An AI company's most valuable asset is a file. Whoever holds a model's weights holds the model, and RAND says stolen weights would run for under half a cent per 1,000 tokens.
  4. Copying a model doesn't need a break-in. Distillation teaches a cheap model to imitate a better one from its answers, and Claude's usage policy bans training any model on its outputs without permission.
  5. Hiring is the third route, and the courts look at files. A judge ordered OpenAI to hand Apple its interview materials for Apple candidates.

The video version, presented by our co-founder Kurt Loy. 7 min 48 sec. Watch on YouTube

We use AI tools every day, and the headline story here, a hidden recorder inside an AI lab, is the least useful part for a business. What matters more is what the evidence is made of, and what the companies building these tools are now doing to protect their models, their reasoning and their staff.

This is the written version of the video above, for teams that use AI at work. We worked from the court record, company statements and press reports as of 16 October 2026, and the sources are at the end.

A note on the case: this is a live criminal matter. Every statement here about the accused is something prosecutors allege, not a proven fact. No court has convicted him, and we are not naming him.

What happened at OpenAI?

Prosecutors allege that an OpenAI engineer hid a small recorder behind the TV in a room OpenAI's executives meet in. The FBI arrested him on 9 October 2026.

According to the criminal complaint, OpenAI staff found a thumb-sized circuit board on 30 September, plugged into a USB port on the back of the TV and wrapped in toilet paper. The complaint says the engineer, who joined OpenAI in July 2025, had no authority to use the room, that security cameras showed him near it on ten days in September, and that the device was installed around 28 September. After the find, a camera was put in the room. Prosecutors say it showed him checking behind the TV on 5 October, and that security staff found a second device that day in a bag he had left in a printer room. They also told the court he had booked one-way flights out of the country for 12 October. His lawyer says the trip was "prepaid".

OpenAI's post about the find has been viewed more than 25 million times. According to a post by the FBI Director, the device was allegedly placed to pick up discussion of the company's "super intelligence tech". That wording comes from a 29 September executive order that tells federal agencies to say "super intelligence" where they used to say "AI", so it tells us nothing about OpenAI's technology.

What is the OpenAI employee charged with?

One count, under the federal wiretap law (18 U.S.C. § 2511), of attempting to use a device to intercept spoken conversations. The maximum sentence is five years.

That is the whole charge. Nobody has been charged with espionage, stealing trade secrets or acting as a foreign agent, and no filing alleges a motive or claims he acted for anyone else. The complaint says none of the files on the first device could be recovered, and OpenAI says it has no indication that he accessed model weights, source code or customer data. His lawyer told the San Francisco Standard that this is a federal recording-law charge, "not stealing trade secrets or working as a foreign agent".

On 16 October the docket listed a detention hearing for 19 October, US time. Federal rules generally require an indictment within 30 days of an arrest, and the charges could change at that point.

Can AI chats be used as evidence?

Yes, and this complaint already relies on them.

Two sets of AI conversations appear in the filing. The first is questions he allegedly asked OpenAI's internal productivity assistant in August and September about the company's meeting rooms, how big they were and when they were booked. The second is prompts about building and concealing a recording device, sent from a chatbot account linked to his spouse, who has not been charged. According to Fortune, OpenAI confirmed the chatbot in question was ChatGPT, and OpenAI says it used its own models to work out what the device was. Nobody has said publicly how investigators got hold of those chats.

It isn't the only recent example. In January 2026 a US judge upheld an order for OpenAI to produce 20 million de-identified ChatGPT logs in a copyright case, one of several data-retention points we covered in our piece on local and cloud AI for teams.

For a business, the useful habit is to treat every prompt as a document someone may read later. Questions typed into a company's own AI tools are visible to that company, as the queries to OpenAI's internal assistant were here. Chats with an outside provider are held under that provider's terms, and in a dispute they can be requested like any other record.

What do AI companies protect most?

The model weights, above everything else. They are the billions of numbers a model learns in training, and whoever holds them can run the model.

The money explains why. OpenAI was valued at $852 billion in March 2026. RAND, the US research organisation, wrote in 2024 that training GPT-4 reportedly ran to $78 million, whereas running a stolen copy of the weights would cost under $0.005 per 1,000 tokens. Around the weights sit the training recipes, the chip and data-centre plans, and the conversations held in rooms like the one in this case.

RAND grades attackers on five levels. It puts insiders on the third, alongside organised cybercrime groups, and says protection strong enough to stop the most capable state hackers, its fifth level, is "currently not possible". Most businesses don't hold anything like model weights, but RAND's ranking applies to them too: the people with access are part of the risk.

What is AI distillation?

Distillation means teaching a cheaper "student" model to imitate a stronger "teacher" by training it on the teacher's answers. It captures a model's abilities without ever getting near its weights.

The method itself is legitimate, and labs routinely distil their own models into smaller, cheaper versions. The fight is over rivals doing it to them through networks of fake accounts. On 12 October Anthropic said new defences had cut what distillers manage to collect from Claude by around 90 to 99%, describing that as early evidence, and that it had banned tens of thousands of accounts, many of them within two hours of the account starting to harvest. It named two China-based labs, Zhipu and Moonshot, and neither has responded publicly to the post. Its September threat report counted about 190 million exchanges across five labs.

Anthropic considers the reasoning a model writes out before answering the most valuable thing to copy. In its test, a model trained on Claude's full reasoning scored 80.4% on a coding benchmark, and one trained only on summaries of that reasoning scored 71.8%. That is one reason the major labs now hide their models' thinking or show a summary instead.

Can your business train a model on Claude's answers?

Not without permission. Anthropic's usage policy bans using Claude's inputs and outputs to train an AI model without prior authorisation from Anthropic, and it applies to any model, not only a rival lab's. A brand-voice model or an in-house assistant fine-tuned on Claude's output falls under the same rule. Before training anything on what an AI tool produced, read that tool's terms.

What do AI companies' defences cost businesses?

Some extra friction, and in places extra money. These measures are aimed at copiers, but ordinary customers meet them too.

If any of your automated workflows ask Claude to print its full reasoning, test them, because blocked requests in that category now cost money.

What can employees take when they change jobs?

What they know, broadly. Files are where the lawsuits start.

Under the federal Defend Trade Secrets Act, a court can't block someone from taking a new job just because of what they know. An injunction needs evidence of threatened misuse, and the cases tend to turn on files, devices and code.

The same week, OpenAI was a defendant in a case about exactly that. Apple alleges that former Apple employees now at OpenAI used its confidential information to help OpenAI develop hardware, to gain insider knowledge when interviewing Apple staff, and to coach new hires on avoiding scrutiny when they left Apple. On 13 October a judge ordered OpenAI to produce forensic images it had collected from devices, every document used to guide or conduct its interviews of Apple candidates, and the private chats of two individuals in the case. The next day, Bloomberg reported, the judge said he was "highly likely" to grant Apple an injunction.

OpenAI says it has no interest in other companies' trade secrets, and argues that the case tries to "cast ordinary job interviews as acts of trade secret misappropriation". It now claims that one of its own former staff copied over 100 files into personal accounts before rejoining Apple. Apple says it found no evidence of that. In an earlier dispute, a judge twice dismissed xAI's lawsuit over the staff OpenAI hired from it, ruling that an employer can ask a candidate about their past work.

For any business hiring from a competitor, the lesson from the order is that interview guides, checklists and scripts can end up in front of a judge. Put a no-files rule in writing for new hires, and keep interviews on skills rather than a rival's projects.

What we'd do now

Our advice for businesses using AI, as of October 2026:

We'll update this page after the detention hearing and if an indictment is filed.

Frequently asked questions

What happened at OpenAI?

Prosecutors allege that an OpenAI engineer concealed a recorder behind a TV in a room used by the company's executives, which staff discovered on 30 September 2026, and that a second device was found in a bag of his on 5 October. The FBI arrested him on 9 October. He has not been convicted, and his lawyer told Fortune the allegations are wholly out of character.

What is the OpenAI employee charged with?

One count under the federal wiretap law, 18 U.S.C. § 2511, for attempting to use a device to intercept spoken conversations, with a maximum of five years in prison. There is no espionage, trade-secret or foreign-agent charge, and no filing alleges a motive. An indictment could change the charges later.

Was anything recorded at OpenAI?

Nothing public shows that it was. The complaint says none of the files on the first device could be recovered, and OpenAI says it has no indication that model weights, source code or customer data were accessed.

Can AI chats be used as evidence?

Yes. The complaint in this case cites queries to OpenAI's internal assistant and prompts from a ChatGPT account linked to the accused's spouse, who is not charged. Separately, a US judge upheld an order in January 2026 for OpenAI to produce 20 million de-identified ChatGPT logs in a copyright case. Treat prompts as records that someone may read.

What are AI model weights?

Weights are the billions of numbers a model learns during training. Whoever holds a copy can run the model, which is why AI companies guard them most closely. A 2024 RAND report puts GPT-4's reported training bill at $78 million and the cost of running stolen weights at under $0.005 per 1,000 tokens.

What is AI distillation?

It means training a low-cost student model on a stronger teacher model's answers so it learns to respond the same way. Labs do it to their own models to make smaller versions, and object when rivals do it through fake accounts. Anthropic reports that new safeguards have reduced what distillers harvest from Claude by an estimated 90 to 99%.

Can I train my own AI model on Claude's outputs?

Not without permission. Anthropic's usage policy bans using Claude's inputs and outputs to train any AI model without prior authorisation, which includes a brand-voice model or an internal assistant. Check each provider's terms before fine-tuning on output from its tools.

Can a company stop an employee from joining a competitor?

Under the US Defend Trade Secrets Act, a court can't block someone's new job just because of what they know; it needs evidence of threatened misuse. Disputes usually turn on files, devices and code. In Apple's case against OpenAI, a judge ordered OpenAI to produce the materials it used to interview Apple candidates.

Sources

We read the court filings, company statements and reports available up to 16 October 2026. The case is ongoing, so check the docket for anything newer.