The Take — The leak risk starts where the protective order ends

A federal judge has given the Justice Department until Friday to explain why a defendant in the OpenAI bugging case should not be allowed to run the evidence through Anthropic's Claude. The government will lose that argument as it has been framed, and it should — because the objection is not really about confidentiality. It is about a protective order with no clause that reaches the machine.
Abdulrahman Alfozan, a former OpenAI engineering lead, has pleaded not guilty to one count of intercepting oral communications, faces up to five years and a $250,000 fine, and has been held without bail since his arrest in October; his trial begins January 20, CNBC reported from Tuesday's hearing. As our brief on that hearing laid out, the defense wants Claude to review discovery, and prosecutors object that one of the seized laptops may hold proprietary OpenAI material. Judge Vince Chhabria was unimpressed — "I'm a lot less concerned about that than you are" — and ordered the government to file a motion by Friday justifying the objection.
I think the judge is right, and I think everyone in the argument is asking the wrong question.
The wrong noun
The government is arguing about the vendor. The risk lives in the pipeline.
Defense teams have run documents through commercial AI for years, and litigation has outsourced review to third-party vendors for two decades — vendors that also work for the other side, that store data in their own clouds, that employ people nobody in the courtroom has met. Courts did not ban those vendors. They imposed vendor-conduct clauses, confidentiality undertakings and clawback orders. The instrument exists. It is not exotic and it is not new.
"Competitor" is a category with no edges. Microsoft sells against OpenAI and hosts its models; Google does both; an unusual share of the AI industry was founded by people who used to work at OpenAI, which is roughly true of Anthropic, a company reportedly days from its own Nasdaq listing. If rival-ness disqualified a tool, the only permissible reviewer would be one the complainant approves — a strange rule to hand a criminal defendant.
Nor does "who built the model" answer "where does the prompt land". The same model can be served through a cloud marketplace, a reseller, or a deployment the customer controls. Only the second question is a confidentiality question.
Where the leak would actually happen
Three places, none of them exotic. Retention: operators keep prompts and outputs for abuse monitoring, for debugging, or because the default says so. Training: consumer tiers and enterprise contracts carry different terms, and the terms change. And the reviewer's habits: the commonest way sealed material escapes a protective order is still a lawyer pasting it into a chat window with no contract attached at all.
Apply that to the specific worry. If the laptop holds OpenAI trade secrets and it is fed to a tool whose operator keeps a copy, the exposure is not that a rival model "learned" something. It is that a copy of the evidence sits in a third party's storage, reachable by subpoena, breach, or an employee who should not have access. That harm is real and irreversible in the way prosecutors say — and it has nothing to do with who trained the weights.
Which makes the useful question not yes or no, but: what does the order say?
What I would put in the order
Five terms, all boring, all enforceable against the party the court actually controls — the defense:
- A written attestation from the operator for this deployment: no training on the inputs, no retention past the session, named personnel only.
- Processing logs preserved and filed under seal, so the court can see what went where without seeing the evidence.
- A segregation protocol for the seized laptop: material that is genuinely OpenAI's goes to human reviewers or a model the defense runs on its own hardware; the rest goes through the tool.
- A clawback-and-destroy provision if anything falls outside the terms.
- A standing invitation for the government to propose a narrower term — because a party that cannot name one is expressing discomfort, not identifying a risk.
Run locally, the problem disappears entirely: an open-weight model on hardware the defense controls cannot leak to anyone, and it is worse at the job. That is a real trade, and the defense should have to justify refusing it. What the defense should not have to do is prove a commercial tool is permissible in principle.
None of this is invention. Federal courts already write Rule 502(d) orders so that producing documents does not waive privilege, and they already require counsel to bind e-discovery vendors to the terms of a protective order. The paper exists. The gap is that nobody has extended it to a model.
The strongest case for the other side
The prosecutor's concern is not paranoid, and it deserves its best form.
The harm is irreversible in a way most discovery disputes are not. If OpenAI's material lands in the logs of a company in registration for a public listing, no sanction un-leaks it, and the defendant — in custody, facing five years — is judgment-proof. A protective order binds the lawyer and the experts; it does not bind a vendor's staff, and it cannot bind weights, wherever they sit. The Sixth Amendment guarantees competent counsel, not the newest tool: human reviewers, or the local model, do the same job. And there is a genuine cost to judges designing AI governance from the bench, one district at a time, in a country whose courts cannot agree on what a chat log is.
I take all of that seriously. Especially the last one.
Why the take still holds
Because the alternative is worse and unadministrable. A rule that lets the government veto a defendant's tooling, in a case where the prosecution's own evidence theory is thin, is a rule applied hardest to defendants least able to litigate it — the public defenders who cannot afford a Friday motion, let alone an appeal.
And the record matters. The government is holding a man without bail on a charge it declined, in open court, to connect to espionage, on a recording device it cannot decrypt, from a phone it cannot open. The detention fight we covered in October turned on the flight, not on the tape. Assistant US Attorney Nick Parker said he was not aware of evidence of economic espionage, trade secrets or unregistered foreign agency, and that he was "not suggesting the defendant stole anything," per The Hacker News. A court that narrows this defendant's ability to prepare, on a risk it cannot articulate as a term, trades a real right for a hypothetical harm.
The precedent is the part that outlives January. Every court in the country is about to face this, and the first order becomes the template. Two bad templates are available: a ban, which quietly becomes a rule that poor defendants get the worst tools; and a shrug, which lets sealed material flow into consumer products because nobody wrote the clause. The good template is dull — terms, logs, clawback — and it is available now for the price of one filing.
What would change my mind
If Anthropic's contract for this deployment permits training on inputs or indefinite retention and it refuses a carve-out, the specific objection is correct and the defense should pick another tool. If the logs show documents went anywhere beyond review, the arrangement is broken. If Friday's filing shows the laptop cannot be segregated without exposing a third party's material — a genuinely hard problem — then restrict the material, not the vendor. And if a court finds no protective order can reach a non-party vendor's systems at all, the answer is a standing order or a rule for appointed counsel: a change that binds everyone at once, instead of a ruling that binds one man on one Friday.
Until then, the sensible outcome is the boring one. Let Claude read the evidence, and make the order say what happens to it afterwards. The confidentiality problem in AI review is not a question of who built the model. It is a question of who wrote the clause — and so far, nobody has.
If an AI tool reviews sealed evidence, should the court publish the terms it approved — or is the order itself something the public has no right to see? Write to us via our contact page.







