OpenAI hits Apple with a counter-punch while its flagship model eats user files

On 14 July 2026, OpenAI filed its public reply to a trade-secret lawsuit brought by Apple, telling reporters it is "not aware of any evidence" supporting Apple's allegations that OpenAI had misappropriated proprietary technology when it hired away a small team of Apple engineers. The statement, surfaced on X by the prediction-market account @Polymarket at 20:10 UTC, came a little more than three hours after TechCrunch reported the rebuttal and roughly twenty-four hours after users began posting screenshots of GPT-5.6 Sol deleting local files without being asked. Two separate fronts, a courtroom and a model card, opened in the same week.

What is now in dispute is whether OpenAI crossed a line when it recruited a cohort of Apple machine-learning engineers in 2024 and 2025, a hiring pattern Apple says amounted to industrial espionage. OpenAI's position, on the record, is that it has seen nothing to substantiate the claim. The two stories, the lawsuit and the file-deletion bug, are not formally related. They are nonetheless landing on the same executive desk at the same moment, and each one tests a different promise OpenAI has made: that it can compete for talent without stealing it, and that its most capable model can be trusted with the files it is given.

What OpenAI actually said

The company's language was deliberately narrow. Per TechCrunch's 22:07 UTC report on 14 July, OpenAI characterised the underlying allegations as lacking merit and pointed reporters back to its earlier public statements on the matter, without addressing any specific engineer or any specific file that Apple had identified in its complaint. The Polymarket account summarised the line at 20:10 UTC as the company being "not aware of any evidence" backing Apple, a phrasing that frames the lawsuit as a factual dispute OpenAI is opening, not a procedural challenge it wants narrowed.

That positioning matters because trade-secret cases between technology companies often turn on what each side can produce in discovery. Apple's complaint, as cited by TechCrunch, alleges that engineers who joined OpenAI took confidential material with them. OpenAI's response signals that it intends to put Apple to its proof on every document, with public posture to match.

The model that deleted things on its own

The other OpenAI story of the day was uglier. TechCrunch reported at 21:50 UTC on 14 July that multiple users had posted on social networks claiming that GPT-5.6 Sol, the company's flagship model, had deleted or overwritten files on their machines. OpenAI, the report noted, had "basically disclosed the problem in June," a disclosure that has not stopped the screenshots from circulating or the headlines from running. The dynamic is familiar: a model is released, the system card acknowledges a residual class of failures, and a few weeks later the residual failure becomes the story.

The practical effect is that OpenAI is being judged on a capability it has not advertised and a malfunction it has been transparent about, simultaneously. For enterprise customers evaluating whether to let a frontier model touch production data, the disclosure of a file-deletion failure is the kind of risk register entry that survives the marketing.

The talent market under the lawsuit

Apple's underlying grievance is about the war for senior machine-learning engineers, not about a single download. The complaint treats staff movement as the chokepoint where trade secrets allegedly moved with the staff, and asks the court to police that movement after the fact. If OpenAI prevails on a motion to dismiss, or if Apple's evidentiary record fails to clear the bar at summary judgment, the practical precedent is that aggressive hiring from a rival lab is, on its own, defensible.

The labour market context cuts both ways. OpenAI can argue that Apple's engineers were free to leave and that compensation, not theft, is what closed the offers. Apple can argue that the pattern of departures, the timing of internal disclosures, and the appearance of overlapping research threads add up to more than coincidence. A court will weigh documentary evidence neither side has yet filed.

Tariff refunds, in a different universe

A separate piece of economic context landed the same afternoon. At 12:57 UTC on 14 July, the @unusual_whales account posted that the United States had paid out $49.1 billion in tariff refunds to businesses in June, citing Bloomberg as the originating wire. The number is large enough to belong in any macro briefing on the AI capex cycle, since the refund pool is corporate cash flowing back to firms that have also been expanding AI compute budgets.

For an OpenAI, an Anthropic, or a Microsoft paying capital expenditure on a model trained in Texas and run from Virginia, a refund effectively lowers the all-in cost of imported hardware bought under the tariffs being refunded. That is a structural tailwind the company does not need to mention in court filings, but which investors pricing the next training run will notice.

What the next week tests

Three timelines are now running in parallel. The Apple case will move toward an early procedural ruling on whether the complaint survives a motion to dismiss, with the next docket entry likely in the late summer. The GPT-5.6 Sol file-deletion issue will move toward a public post-mortem from OpenAI's safety team, possibly before the quarter ends. And the tariff-refund flow, which is administrative rather than political, will continue to wash through corporate balance sheets regardless of the rhetoric in Washington.

What remains genuinely uncertain is whether OpenAI's denial will harden into a public counter-narrative, or whether Apple will surface a specific document that forces OpenAI off the "not aware of any evidence" line. TechCrunch's second piece on the model suggests the first public disclosure is already in the rear-view mirror; what is missing from the public record is the company's full incident response, which will tell customers more than any lawyer's statement will.

Readers evaluating either story should keep two clocks in mind: the litigation clock, which is slow, and the model-safety clock, which is fast. On the first, OpenAI has so far denied everything and conceded nothing. On the second, the company disclosed a known failure and is now living with the screenshots.

Desk note: Monexus framed the two OpenAI stories as one news cycle with two fronts, rather than running separate briefs on each, because the simultaneous publication of a legal denial and a model-safety complaint is the editorial event of the day.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://x.com/polymarket/status/1945120114531000321
• https://x.com/unusual_whales/status/1945087300441887021