OpenAI Apple Trade Secrets Deal Puts 14 Employee Devices Under Review

OpenAI agreed to restrict its use of disputed Apple information and preserve forensic copies from devices used by 14 employees. The OpenAI Apple trade secrets deal stops short of admitting that anything was stolen. Yet it gives a federal judge authority to enforce restrictions while the companies fight over hardware development, recruiting, and confidential data.

The voluntary agreement followed a hearing on October 14, 2026, before U.S. District Judge Edward Davila in San Jose. Davila reportedly said he was “highly likely” to grant Apple preliminary relief before OpenAI made its concession. The judge can enforce the arrangement during the litigation, according to a court hearing account.

That sequence creates the central tension. OpenAI denies needing Apple’s confidential information and says it is creating a different class of AI devices. Apple argues that former employees carried protected hardware knowledge into a business that could eventually compete with its most important products.

The agreement does not resolve those competing claims. It instead converts them into an evidence-preservation and compliance problem. OpenAI must now separate contested information from legitimate employee knowledge while continuing a hardware program built partly by former Apple talent.

What the OpenAI Apple Trade Secrets Deal Changes

OpenAI has accepted an enforceable operational restriction without conceding that Apple’s allegations are true.

The agreement reportedly bars OpenAI from accessing, using, or disclosing trade secrets that Apple claims were misappropriated while the case proceeds. That wording matters. It describes the information as Apple’s alleged trade secrets rather than established stolen property.

A trade secret is commercially valuable information that remains protected through reasonable efforts to keep it confidential. Courts do not treat every internal document or employee memory as a trade secret. The owner generally must identify the information and show that it was protected from unauthorized disclosure.

Apple still carries that burden. OpenAI can maintain that Apple has defined its protected information too broadly, that some information was already known, or that its employees developed relevant work independently. The interim promise does not decide any of those questions.

However, the promise has practical force because Davila can enforce it. If Apple presents credible evidence that OpenAI used covered information after making the commitment, the dispute could expand beyond the underlying allegations. The court could examine whether OpenAI violated an enforceable agreement and consider sanctions or stronger restrictions.

OpenAI also agreed to collect forensic images from devices used by 14 employees whose equipment might contain Apple information. A forensic image is a preserved copy of a device’s data, including files, metadata, logs, and recoverable deleted material.

Creating an image does not prove that its owner stole or used anything. It freezes evidence so investigators can reconstruct what happened without repeatedly handling the original device. That distinction protects both sides because the resulting record can support or undermine Apple’s account.

The 14-device commitment is broader than an investigation limited to the two former Apple employees named as defendants. Apple sued OpenAI alongside Chang Liu, Tang Yew Tan, and io Products. Liu worked as an electrical engineer at Apple, while Tan spent 24 years there and later became OpenAI’s chief hardware officer.

Apple subsequently said its investigation identified 11 additional former employees who might have relevant evidence or involvement. Those individuals were not necessarily accused of theft. Their devices could contain communications, transferred files, interview materials, or evidence showing that no improper transfer occurred.

The collection requirement also differs from unrestricted access to every OpenAI system. Apple had sought broad information about repositories and devices that might contain what it called Apple-originated information. Davila expressed concern that some of Apple’s requested categories were too expansive.

This limitation is significant. The court has preserved evidence from an identified group without immediately opening OpenAI’s entire hardware operation to Apple. OpenAI therefore faces a substantial compliance burden, but Apple has not received unlimited access to a competitor’s research.

The agreement follows a separate October 13 order partly granting Apple expedited discovery. Davila ordered the defendants to respond quickly to requests involving forensic images already collected, recruiting materials, selected communications, and employee responsibilities.

Materials already gathered for the litigation must be produced on a shorter schedule. Other covered materials receive a longer deadline. The underlying discovery order gives Apple evidence before the case reaches the ordinary discovery timetable.

That combination changes the immediate balance. Apple gains preserved evidence and targeted internal records. OpenAI retains room to contest Apple’s definitions, challenge the lawsuit, and protect research that the judge did not place on an expedited path.

Why Fourteen Forensic Images Matter

The device images turn a dispute about corporate intent into a test of specific files, timestamps, messages, and user actions.

Apple’s complaint presents the alleged misconduct as a coordinated pattern. It claims former employees accessed confidential files, discussed protected projects during recruitment, and carried knowledge into OpenAI’s consumer hardware work. OpenAI disputes that characterization.

Forensic evidence can narrow that gap. Investigators can examine when a file arrived, whether someone opened it, where it was copied, and which account transferred it. They can also determine whether an allegedly sensitive file remained untouched or was deleted before an employee began work at OpenAI.

Metadata rarely answers every question by itself. A file appearing on a device does not establish that the employee understood its contents, shared it with colleagues, or incorporated it into a product. A missing file also does not prove that the file never existed.

Context must come from messages, interviews, design records, access logs, and testimony. That is why the October 13 order covers more than device images. It also reaches certain recruiting records and LINE messages involving Liu and former Apple colleague Alyssa Peng.

Apple alleges that Liu retained an Apple-issued computer and accessed confidential hardware materials after leaving. The company also claims that he discussed Apple information with people seeking work at OpenAI. Liu and OpenAI contest Apple’s interpretation of the evidence.

Tan occupies a different role in Apple’s theory. Apple alleges that he used confidential project language when interviewing candidates and encouraged applicants to discuss components or work from unreleased products. OpenAI argues that Apple has recast normal technical recruiting as trade secret solicitation.

Hiring experienced engineers inevitably creates difficult boundaries. An employee can use general skill, professional judgment, and knowledge developed over a career. The employee cannot take protected documents or disclose specific confidential designs belonging to a former employer.

The distinction becomes harder when a senior hire has spent decades working on tightly integrated products. Expertise in materials, manufacturing, batteries, power management, and industrial design does not disappear when employment ends. Yet specific unreleased specifications can remain protected.

The OpenAI Apple trade secrets dispute therefore concerns both files and organizational controls. Investigators will look at what employees brought with them, but they will also examine how OpenAI handled that material once concerns appeared.

Useful questions include whether OpenAI quarantined questionable files, notified Apple, restricted affected employees, and compared disputed material with its product records. The company’s ability to document those steps will influence how credible its independence argument appears.

OpenAI’s recruiting procedures face similar scrutiny. Apple has sought interview guides, instructions, scripts, and technical presentations used with candidates from its hardware teams. Those materials could show whether interviewers asked candidates to solve general engineering problems or reveal protected Apple work.

Davila gave Apple expedited access to selected hiring records. He also authorized early depositions of Liu, Peng, and Tan. A deposition is sworn testimony taken before trial, allowing lawyers to question witnesses while memories and electronic evidence remain relatively fresh.

The judge did not grant everything Apple requested. He declined to expedite discovery into areas including metal finishing, power management, and battery functions. He found that those materials could proceed through the ordinary schedule.

That restraint cuts against treating Apple’s allegations as proven. It suggests that the court sees enough urgency to preserve defined evidence but remains unwilling to expose broad sections of OpenAI research without narrower justification.

For OpenAI, the 14 forensic collections introduce technical and managerial costs. The company must preserve potentially relevant devices, separate personal or unrelated corporate information, coordinate experts, and maintain a defensible chain of custody.

For Apple, the images create risk as well as opportunity. If the evidence does not show meaningful use or dissemination, its account of a coordinated campaign could weaken. Targeted digital records are harder to reconcile with sweeping allegations when the records point elsewhere.

This is why forensic preservation is more than courtroom housekeeping. It establishes a common body of evidence against which both companies’ public claims can be tested.

Apple and OpenAI Are Fighting Over More Than Files

The primary conflict is whether OpenAI built its hardware organization through legitimate hiring or through access to Apple’s protected product knowledge.

Apple filed its lawsuit in July in the U.S. District Court for the Northern District of California. It alleged that OpenAI used former Apple employees, recruiting interviews, and supplier relationships to accelerate its move into consumer hardware.

The company’s complaint names OpenAI, Liu, Tan, and io Products. Apple says Liu downloaded confidential hardware files and that Tan sought protected information from employees considering jobs at OpenAI. The defendants have denied the central allegations.

OpenAI acquired io Products, the hardware company associated with former Apple design chief Jony Ive, in a deal valued at nearly $6.5 billion. Ive is not named as a defendant, but the acquisition places former Apple leadership and design experience near the center of OpenAI’s hardware strategy.

That strategy increases the stakes. OpenAI is not merely hiring engineers for software infrastructure. It is developing a family of devices intended for an AI-centered computing model, according to its court filings and public statements.

Apple says those plans create a direct risk that its confidential engineering work will become embedded in competing products. Davila’s discovery order recognized the concern that disputed information might become so intertwined with OpenAI research that later remedies would be inadequate.

OpenAI presents a different story. In its motion to dismiss, the company said it had “no use, need or desire” for Apple trade secrets because it was building something different. It also argued that Apple had failed to describe the protected information with sufficient precision, according to the dismissal filing coverage.

This disagreement places employee mobility at the center of the case. Apple reportedly claims that more than 400 former employees have joined OpenAI. The number illustrates the scale of talent movement, but it does not establish wrongdoing by those workers.

Silicon Valley companies routinely recruit from competitors. California policy generally favors employee mobility, while federal and state law still protect qualifying trade secrets. Companies must therefore distinguish portable expertise from proprietary information.

Apple’s strongest theory depends on concrete transfers or solicitations. Its allegations about retained equipment, downloaded documents, internal project names, and physical components seek to cross that line. OpenAI’s defense depends on separating those alleged acts from its institutional conduct.

The companies also remain business partners. Apple integrated ChatGPT into its software as an optional external model for requests that its own systems could not answer. That relationship makes the litigation more unusual than a conventional fight between isolated competitors.

Their incentives now overlap and diverge. Apple can benefit from offering users access to OpenAI models while resisting a new device platform that challenges the iPhone’s role. OpenAI can benefit from Apple distribution while trying to reduce its dependence on smartphones and app stores.

The original lawsuit represented a sharp turn in that relationship. Apple’s filing called the matter an effort to stop former employees from using its secrets for OpenAI’s benefit. OpenAI initially said it had no interest in another company’s protected material.

The broader lawsuit allegations include claims about unannounced technologies, engineering documents, manufacturing processes, and supplier information. Every category remains an allegation unless established through evidence or adjudication.

OpenAI’s agreement does not concede that its products contain Apple work. It does, however, require the company to operate as though disputed information creates a serious contamination risk.

That is an important reversal. OpenAI can continue arguing that Apple’s lawsuit is baseless, but it must simultaneously identify, preserve, and avoid information covered by Apple’s claims. The practical burden arrives before the legal merits are settled.

For other technology companies, the lesson is not that hiring from competitors has become unlawful. The lesson is that employers need documented procedures for handling confidential information carried by new hires.

Those procedures can include written instructions, device screening, file quarantine, clean-room development, and restrictions on using former-employer materials. A clean room separates a project team from disputed information so the company can document independent development.

The need becomes more urgent when recruiting happens at scale. A company can tell every new hire not to bring files, yet informal interview practices or weak onboarding controls can undermine that policy. Courts will examine conduct, not only written rules.

OpenAI’s Counterclaim Threat Complicates Apple’s Story

OpenAI now alleges that confidential hardware information traveled in the opposite direction, turning a one-way theft narrative into competing unresolved accusations.

Hours before the October 14 hearing, OpenAI alleged that a former employee moved more than 100 confidential files to personal accounts before returning to Apple. The person had previously worked for Apple and spent six months at OpenAI, according to OpenAI’s filing.

OpenAI said the files concerned at least one unreleased device and included images, engineering information, manufacturing data, supplier details, and finishing standards. It also alleged that the employee presented information about OpenAI’s plans to Apple leadership.

Apple told reporters it had found no evidence supporting those claims. The company said it maintains policies intended to prevent new employees from using confidential information belonging to former employers.

Neither account has been proven. OpenAI’s filing does not erase Apple’s allegations, just as Apple’s lawsuit does not establish that OpenAI misappropriated anything. The two claims require separate evidence and legal analysis.

Still, the counter-allegation changes the strategic picture. OpenAI can argue that confidential-data risk follows employee movement in both directions. It can also ask whether Apple applied to the returning employee the controls it demands from OpenAI.

The reported transfer date was June 2, 2026, the employee’s final day at OpenAI. Specific timestamps and file counts make the claim testable. Investigators can examine account records, device logs, presentation materials, and the employee’s communications.

OpenAI said its forensic investigation uncovered the activity during this lawsuit. That timing shows how evidence collection can generate claims unrelated to the original theory. Once a company begins imaging devices and tracing transfers, the investigation can expose its own vulnerabilities.

The counterclaim filing also illustrates why Apple’s request for broad access created concern. Each side possesses sensitive research that could reveal unreleased products, suppliers, and design decisions to a developing competitor.

Courts can manage that risk through protective orders, outside experts, access limits, and sealed filings. Even with those safeguards, discovery involving unreleased hardware requires careful boundaries.

The competing claims do not create a simple equivalence. Apple alleges an institutional pattern involving recruiting and multiple former employees. OpenAI’s new allegation concerns one unnamed former employee and a potential future claim.

Evidence might validate one allegation, both, or neither. Readers should resist assuming that misconduct in one direction would excuse misconduct in the other. Trade secret duties apply independently to each company and employee.

A second uncertainty concerns the scope of OpenAI’s promise. Apple and OpenAI still need workable criteria for identifying information subject to the restriction. Apple’s broad “Apple-originated information” concept has already drawn judicial concern.

If covered material is defined too narrowly, Apple may argue that the agreement offers little protection. If it is defined too broadly, OpenAI could struggle to distinguish protected information from general engineering knowledge held by former Apple employees.

That ambiguity creates compliance risk. An engineer might recognize a common manufacturing problem through years of experience without using a particular Apple document. Another engineer might unknowingly receive a design influenced by material that moved through several colleagues.

The court will need evidence connecting information, access, and product decisions. Similarity between two designs alone might not show copying. A file transfer alone might not demonstrate product use. The persuasive case will likely combine technical records with human communications.

A third uncertainty concerns OpenAI’s hardware schedule. A preliminary restriction can cause delay even when it does not formally halt development. Teams might repeat analyses, isolate employees, or review design history to reduce contamination concerns.

Apple sought broader expedited access to OpenAI’s hardware research, but Davila declined that request at this stage. That gives OpenAI room to continue development while preserving the possibility of ordinary discovery later.

The agreement therefore represents a tradeoff for both companies. Apple gains enforceable protection and access to targeted evidence. OpenAI avoids an immediate court-drafted injunction that might have imposed less predictable restrictions.

Three Signals Will Decide What Happens Next

The next phase will be shaped by the agreement’s written scope, the forensic evidence, and any connection between disputed material and OpenAI’s device work.

The first signal is the final language used to implement the voluntary commitment. A public order or stipulation should clarify who is covered, what conduct is prohibited, and how disputes over contested information will reach the judge.

Definitions will matter more than dramatic courtroom statements. The phrase “Apple trade secrets” sounds precise, but the parties disagree over which information qualifies. Enforcement becomes difficult if employees cannot determine what falls inside the restriction.

A narrow definition tied to identified files would improve operational clarity but might leave Apple concerned about related knowledge. A broad definition tied to entire product areas could interfere with legitimate work by employees who gained experience at Apple.

The written process for forensic review is equally important. The parties must decide who creates the images, who can search them, which terms or time ranges apply, and how privileged or unrelated information will be protected.

The second signal is what the 14 device images actually show. Readers should watch for evidence of file movement, access after employment changes, sharing among colleagues, and links to recruiting interviews.

Absence will be informative too. If investigators find that alleged files were never opened at OpenAI or never reached relevant teams, the company’s independent-development defense becomes stronger. If records show repeated transfers and discussion, Apple’s request for broader discovery gains support.

The October 13 ruling already reflects concern that logs and metadata can disappear through ordinary system use. That is why preservation happened early. The judge did not need to find deliberate destruction before protecting potentially short-lived evidence.

Davila partly granted expedited discovery while rejecting wider requests. The early discovery ruling therefore provides a useful baseline. Future expansion will likely depend on whether initial evidence justifies moving into OpenAI’s research files.

The third signal is any documented connection between Apple information and OpenAI product decisions. This is the hardest and most consequential question.

Apple will seek evidence that disputed material influenced specifications, manufacturing choices, supplier discussions, or design milestones. OpenAI will seek contemporaneous records showing that its teams reached those decisions through independent research.

Version histories, meeting notes, simulation records, supplier communications, and dated prototypes can establish development paths. Well-kept technical documentation becomes legal evidence when a company must show where an idea came from.

That is relevant beyond this lawsuit. AI companies are moving into hardware while recruiting people from established device makers, semiconductor companies, laboratories, and robotics teams. Each move carries knowledge that is valuable precisely because it came from earlier work.

Companies cannot erase an employee’s experience. They also cannot treat a new hire’s previous employer as a free research archive. The practical boundary depends on disciplined information handling and evidence of independent development.

For developers and technical leaders, the case offers a concrete governance lesson. Preserve design decisions, record source materials, and make confidentiality reviews part of onboarding. Do not wait for a lawsuit before separating personal accounts, former-employer files, and company systems.

Knowledge workers should also understand that forensic imaging can preserve more than visible documents. Metadata, synchronized folders, messaging records, and deleted-file remnants can reconstruct actions that seemed temporary.

The OpenAI Apple trade secrets fight is not resolved by the October agreement. It has simply moved from accusations into a controlled evidentiary phase, where both companies must support their stories with records.

Watch the written restriction first, the forensic findings second, and any product-development link third. Those signals will show whether the agreement remains a temporary safeguard or becomes the foundation for a broader injunction.

OpenAI’s hardware program can continue, but it now operates under judicially enforceable constraints and a 14-device evidence review. Apple has gained leverage without yet proving misappropriation. The decisive question is no longer whether the companies can trade accusations. It is whether either side can trace confidential information into real engineering decisions, and what that record says about how the OpenAI Apple trade secrets dispute should end.