OpenAI Promised Not to Use Apple's Secrets. The Promise Is the Tell.

Companies that are sure they're innocent rarely volunteer to stop doing the thing they've been accused of. On Wednesday in a San Jose courtroom, OpenAI did exactly that. It agreed to leave alone any Apple trade secrets its hires may have brought with them while the case runs, and to collect forensic copies of 14 employees' devices that may hold Apple material. It did so after U.S. District Judge Edward Davila said he was "highly likely" to grant Apple an injunction anyway.

When Apple filed this suit in July, we argued it fit a pattern: the fourth time OpenAI had broken a partner's trust. Three months later, the case has moved from accusation to the first real signal from the bench. That signal mostly favors Apple. The more interesting question is what each company is actually fighting for, because it isn't really a few hundred files.

How we got here

The relationship looked very different two years ago. In June 2024, Apple and OpenAI announced the ChatGPT integration that still ships inside iOS today. The turn came in May 2025, when OpenAI bought Jony Ive's io for $6.4 billion and made clear it intended to build its own consumer hardware. Building hardware means hiring people who know how to build it, and the deepest bench of those people in the world sits in Cupertino.

The key dates since then:

• July 2026: Apple sues OpenAI, io Products, OpenAI hardware chief Tang Tan and former iPhone hardware engineer Chang Liu in federal court in San Jose (Apple v. Liu, 5:26-cv-07078). The complaint alleges a coordinated effort to pull confidential design, manufacturing, component, testing and supply chain information out of Apple employees, including "show and tell" interviews where candidates were asked to bring prototypes and parts.
• August 3: Apple asks for a preliminary injunction. OpenAI answers publicly with a post titled "Apple is getting this wrong."
• August 5: OpenAI moves to dismiss, arguing Apple never names a specific trade secret, only broad categories of hardware knowledge.
• August 31: Apple tells the court OpenAI is destroying evidence. OpenAI calls the dispute a mess of Apple's own making.
• October 13: Davila grants Apple most of its early discovery: forensic images, the checklists and scripts OpenAI used to interview Apple candidates, interview summaries, and Liu's LINE messages with a former colleague since January.
• October 14: The injunction hearing, OpenAI's voluntary standstill, and OpenAI's own accusation against Apple.

What the judge actually said

The headline is "highly likely," but the details are where Apple's win has limits. Davila called a forensic inspection of every OpenAI device and repository that might hold Apple information a "huge ask" and declined to grant all of it. In Tuesday's discovery order he called Apple's catch all category, "Apple Originated Information," "amorphous as currently defined" and gave Apple five days to tighten it. Research files on metal finishes, power management and battery behavior stay in ordinary discovery rather than being handed over early.

Apple's lawyer, John Desmarais, seemed to see that coming and narrowed the target in the room. "They don't have 400 bad apples, your honor, they have a few bad apples and know who they are," he said, offering to limit inspections to a small group of people who moved to OpenAI.

That line matters more than it looks. Apple is no longer arguing that OpenAI's whole hardware team is contaminated. It's arguing that a handful of specific people did specific things, which is a case Apple can actually prove and a judge can actually enforce.

OpenAI's counterpunch

OpenAI arrived with its own accusation. In a filing the same day, it said a former employee spent about six months at OpenAI, then moved more than 100 confidential files about an unreleased OpenAI device to personal accounts, some of them just after midnight on June 2, 2026, his last day, before returning to Apple. OpenAI says it understands he later briefed Apple leadership on its product plans, and it points to Apple's newly announced HomeView display as a sign the information was used.

Apple says it found no evidence for any of this and calls it a distraction. OpenAI hasn't filed counterclaims; it asked for time to investigate first. And the HomeView link is the weakest part of the story. Apple opened HomeView preorders this week, about four months after that employee walked back in the door. Nobody designs, sources, certifies and ships a $299 smart display in four months.

As a legal defense, "you did it too" doesn't answer Apple's claims. As leverage, it's useful. It gives OpenAI a reason to demand discovery from Apple, a talking point for a company preparing to go public at a reported $1.4 trillion valuation, and a story that makes this look like two rivals poaching each other rather than one taking from the other.

The part neither side wants to dwell on

Strip away the headlines and the most revealing facts in this case are about offboarding, not espionage.

Apple's core allegation is that Liu used an authentication flaw to get back into Apple's network storage after he left and download files. OpenAI's defense leans on the same weak spot from a different angle: it says Apple encouraged staff to use personal iCloud accounts for work, that an authentication error left Liu with access after his departure, and that Apple colleagues kept asking him to dig up files and answer technical questions after he'd gone. In August OpenAI posted Liu's private messages on its website to make that case.

Anyone who has run identity and device management at a large company reads that and winces. Deprovisioning is supposed to be the boring part: the account is disabled on the last day, the tokens are revoked, the laptop comes back, and work data never lived in a personal cloud account to begin with. Whatever Davila decides about OpenAI's conduct, both companies' briefs describe a former engineer who could still reach Apple data and Apple colleagues who treated him as if he still worked there. That's not a hiring problem. It's an access problem, and it belongs to Apple.

It also explains why Apple is fighting this hard. California voids noncompete agreements, so Apple can't stop people from leaving for OpenAI. It can only make an example of what they take with them. A court order, even a narrow one, tells the next wave of recruits and recruiters exactly where the line is and what crossing it costs.

What happens next

The near term deadlines are concrete. Forensic images from the first group of devices are due around October 18, OpenAI's answers to Apple's expedited requests are due within 14 days of Tuesday's order, and depositions follow within 30 days after Apple gets those answers. No date has been set for the injunction ruling or for a decision on OpenAI's motion to dismiss, but Davila's comments make dismissal look unlikely.

For OpenAI, the real cost is time. Its first device, reportedly a doughnut shaped smart speaker priced above $300, is aimed at 2027, and Bloomberg has reported the case has already slowed that work. Every engineer who has to be walled off from a project, and every design decision that has to be documented as clean, pushes that date out a little further, right as the company is trying to show public market investors it can be more than a chatbot.

For Apple, the case has already done much of its job. OpenAI is under a court enforceable promise not to use what Apple says it took, 14 people's devices are being imaged, and the interview scripts that allegedly asked Apple engineers to bring parts to a job interview are headed to Apple's lawyers. And through all of it, ChatGPT is still built into the iPhone. The two companies are still partners on one product while suing each other over the next one, and whichever one wins the hardware fight, that awkward arrangement probably won't survive it.