Apple filed what it calls “shocking evidence” against a former engineer. Read it against the case a San Francisco judge threw out in June and every piece lines up with a hole she identified.
Apple filed new evidence on August 31 in its trade-secret case against OpenAI, built from a MacBook that a former employee’s lawyer handed over earlier in the month. The filing reads as an attack on one engineer. It works as an attack on a ruling. Ten weeks before Apple sued, the same federal court dismissed a nearly identical claim against OpenAI with prejudice, and the judge listed exactly what a plaintiff would need to plead to survive. Apple’s new allegations answer that list item by item. Behind the legal choreography sits a clock: OpenAI has told a court its first device ships no earlier than the end of February 2027, which gives Apple 233 days from its complaint, against a prior case that took 264 days to die before discovery even opened.
You do not usually get to watch a company write a pleading against a precedent in public. Apple’s lawyers had a roadmap published on June 15, they filed 25 days later, and on Sunday they showed the court what they found on the laptop.
What Happened
Counsel for Chang Liu, a former Apple engineer who now works at OpenAI, turned over Liu’s old Apple work laptop in early August. Apple’s August 31 filing describes what its investigators found and calls it shocking evidence that its trade secrets are in use and that evidence is being destroyed.
Apple makes four claims in the filing. Liu used a confidential Apple circuit schematic in his work at OpenAI. Liu worked with a tool that shares a name with an internal Apple engineering application. OpenAI was, in Apple’s words, well aware of Liu’s continued access to Apple data. And in June, when Liu learned Apple was investigating him, he brought in an OpenAI colleague, Yu-Ting Peng, to help destroy evidence.
The underlying exhibits are redacted. Apple frames the laptop as the limited slice of material the defendants produced after weeks of delay, and argues it proves the company is not fishing. Apple wants a preliminary injunction blocking OpenAI from working on hardware derived from its technology while the case runs, plus expedited discovery, on the theory that more former employees are involved.
OpenAI’s public position has not moved. In a blog post this month the company said Liu opened Apple files after leaving only because former colleagues asked him for help, and that residual access is a recurring Apple problem caused by Apple failing to cut off system access when people go. Apple’s answer is that Liu kept his access by exploiting a rare authentication bug nobody knew about.
The Backstory
Elon Musk’s xAI ran this play first and lost it twice.
xAI sued engineer Xuechen Li on August 28, 2025, and Judge Rita Lin ordered Li to hand over his personal devices on September 2, barring him from working on generative AI at OpenAI. On September 24, xAI sued OpenAI itself in the Northern District of California, case number 3:25-cv-08133, alleging that recruiter Tifa Chen coordinated a campaign that pulled eight engineers and executives out of xAI and pulled source code out with them.
Lin dismissed the complaint on February 24, 2026, with leave to amend. Her sentence is the one that matters: notably absent are allegations about the conduct of OpenAI itself. She gave xAI until March 17 to fix it. xAI could not. On June 15 she dismissed the case with prejudice, ruling that mere possession of trade secrets does not amount to misappropriation and that asking a job candidate about prior work is a routine recruiting question rather than inducement.
California made that outcome likely. The state voids non-compete agreements and rejects the inevitable disclosure doctrine, so a departing engineer’s knowledge travels with them by right. To reach the new employer you have to show what that employer did. Our July analysis of Apple’s original complaint covered the same ground from the talent side: Apple’s hardware moat was never a legally protected asset.
One detail from the xAI file explains a lot about how Apple is behaving now. OpenAI extended Li an offer on July 28, 2025, Li accepted on August 1, and OpenAI revoked it once the allegations surfaced. Li never worked at OpenAI. xAI’s fastest, most satisfying win, the September restraining order, removed the only fact that could have proved OpenAI used anything.
The Plan
Chang Liu still works at OpenAI. That single difference is the foundation of Apple’s case, and Apple’s new filing spends its ammunition converting it into the elements xAI never had.
Line the two documents up. Lin wanted allegations about OpenAI’s conduct, so Apple names a second OpenAI employee, Peng, participating in destruction on OpenAI premises. Lin ruled that possession is not misappropriation, so Apple alleges use, a specific circuit schematic applied inside Liu’s OpenAI work. Lin wanted facts showing the defendant’s products or internal systems reflect the stolen material, so Apple points at an OpenAI tool carrying the name of an internal Apple application. Lin wanted knowledge, so Apple states that OpenAI was well aware. Four holes, four patches.
The expedited discovery request follows from the same reading. A trade-secret plaintiff in this district gets nowhere on inference, and Apple is asking the court for permission to go looking early rather than plead its way past a motion to dismiss on characterization alone.
Then there is the arithmetic nobody has run.
xAI’s case against OpenAI ran 264 days from complaint to final dismissal without reaching discovery. Apple filed on July 10, 2026. OpenAI told a different court in February, in the iyO trademark dispute, that its first hardware device would not ship to customers before the end of February 2027. That leaves 233 days between Apple’s complaint and the earliest legal ship date, a shorter window than the last plaintiff needed to lose. Apple cannot win this on the merits before the product exists. The preliminary injunction is not a tactic attached to the claim. It is the claim.
The Business Model Angle
California prices engineer mobility at zero and prices the acquiring employer’s conduct at everything. For a business, that flips where protection money should go.
Start with what Apple’s two strongest allegations have in common. A circuit schematic is identifiable. A tool name is a fingerprint that survived a copy. Neither is an NDA, a confidentiality clause, or an exit interview. Both are artifacts that let Apple point at something inside OpenAI’s building and say that came from us. Trade-secret law in this state rewards the company that can produce that pointer and gives nothing to the company that can only describe how valuable its knowledge was. If you run a business whose value sits in accumulated process, the spend that pays is identification: version stamps, distinctive internal naming, artifact fingerprints, access logs granular enough to tie a file to a person and a date. The Apple business model runs on manufacturing know-how that no patent covers, and this case is the first time Apple has had to prove that know-how exists in a form a court can see.
The second lesson is harder for a legal department to accept. xAI’s restraining order against Li was a clean, fast victory that destroyed the durable claim, because a man kept out of OpenAI can never be shown using anything at OpenAI. Apple let Liu keep working, waited, and went after the hardware. The remedy you can actually win requires the harm to happen first. Enforcement here is retrospective by design, and the instinct to stop the departure at the door forfeits the only case worth bringing.
For the company doing the hiring, the exposure sits in a place most onboarding processes ignore. OpenAI’s defense on the front end is strong and matches the law: recruiters asked routine questions, the company revoked Li’s offer, hiring rivals’ engineers is lawful. Peng is the problem. A second employee allegedly helping with deletion after a preservation obligation attaches is company conduct, and it is the one category of fact that Lin’s ruling says can attach corporate liability. Any firm hiring aggressively from a competitor is carrying a liability that its recruiting policy cannot touch, because the risk is created after the hire, by people responding to a legal threat without supervision. OpenAI’s own cost structure makes hardware its most plausible route out of pure compute economics, which means the device program is worth defending and worth attacking.
Both sides also agree on something neither wants to emphasize. OpenAI says Apple failed to manage system access at offboarding. Apple says Liu exploited an authentication bug. Under either story, the perimeter that failed was technical, not contractual. The OpenAI business model and Apple’s both rest on knowledge nobody registered anywhere, and the only wall around it is identity and access management.
The Risk
Apple’s exhibits are redacted, so the reading above rests on how Apple characterizes its own evidence in a brief written to persuade a judge. Take the confidence level accordingly.
Lin’s June ruling binds nobody in Apple’s case. A different judge can weigh the same doctrine differently, and Apple’s posture is stronger than xAI’s in ways that have nothing to do with clever drafting: its engineer actually works at OpenAI, it raised the issue with OpenAI privately in February before filing, and the Apple value chain produces the kind of concrete hardware artifacts that make a misappropriation claim easier to picture than model training methods.
The stronger counterargument runs against Apple. Trade-secret protection requires reasonable measures to keep the information secret. Apple has now told a federal court that a former employee held live credentials to confidential engineering files for months through a bug in its own authentication system. OpenAI will argue that is not reasonable measures, and every additional detail Apple files about how the access persisted strengthens that defense. Apple is also bound to identify its secrets with particularity to proceed, which means describing them on a public docket to a competitor that hired 400 of its former staff.
Injunctions in trade-secret cases are difficult to obtain and usually narrow when granted. On spoliation, Federal Rule 37(e) requires intent to deprive before a court will impose an adverse inference or terminating sanctions, and deletion by an employee panicking about an investigation does not automatically clear that bar. Peng’s own account is not yet public.
Watch two markers. If the court denies expedited discovery, Apple’s case slows to a normal schedule and resolves after the device ships, at which point it converts into a damages claim about a product already in stores. If OpenAI’s response makes offboarding failure its lead argument rather than a secondary one, that signals its lawyers think the reasonable-measures element is where Apple is weakest.
Quick Questions
Is it illegal for OpenAI to hire 400 former Apple employees? No. California voids non-compete agreements and treats employee mobility as protected. Apple’s claims concern conduct around the hiring, not the hiring.
Why does the xAI case matter to Apple’s? Same court, same defendant, same theory. Judge Rita Lin dismissed xAI’s claims with prejudice on June 15, 2026, ruling that xAI pleaded facts about departing employees rather than about OpenAI. Apple filed 25 days later.
What is Apple asking for? A preliminary injunction blocking OpenAI from developing hardware based on Apple technology while the case proceeds, plus expedited discovery. Damages are secondary to stopping the clock.
Can Apple lose even if Liu did what Apple says? Yes. Liability for the employee and liability for the employer are separate questions, and Lin’s ruling holds that possession by a new hire does not make the new employer a misappropriator.
Does OpenAI’s residual-access defense help or hurt it? Both. It weakens the theft framing and strengthens Apple’s problem on the reasonable-measures element, since it argues Apple never closed the door.
The Business Model Analyst Take
Apple is running a better version of a case that already failed, and the improvement is procedural rather than moral. Its lawyers read a dismissal order, waited for the facts that order required, and filed when a laptop produced them. Companies rarely get that clean a signal about what a court wants to hear.
The lesson for anyone whose business runs on knowledge sitting in employees’ heads: your protection is not the agreement they signed. It is whether you can identify your work inside somebody else’s building, and whether your access controls end when employment does. Apple built the most valuable hardware organization in history on know-how it never patented, and it is now asking a judge to protect that know-how using a circuit schematic and a tool name, because those are the only two things it can point at.
The 233 days are the real story. Trade-secret litigation moves slower than a product cycle, which means the remedy that matters is the one you get in the first year or not at all. Apple knows this. That is why the filing landed on a Sunday night in August, and why it reads less like an accusation against an engineer than a set of answers to questions a judge asked somebody else in June.
