Apple's supplemental brief was filed in support of expedited discovery, and the language its attorneys chose does the work that request needs: learning from trade secret information "may create irreversible and continually propagating uses of the trade secret," harm that is "uniquely challenging to undo and requires prompt investigation" [1][3]. That framing is a procedural move as much as a technical one, and its first payoff shows up on the calendar rather than in a damages column.
The technical fork underneath it is where operators should look. Sijia Liu, a Michigan State computer science professor who co-authored work on machine unlearning, told Business Insider that if a sensitive document sits in a repository an AI system retrieves from, the remedy can be as simple as deleting the file [6]. If the same information was used to train or fine-tune a model, the process is different and likely resource-intensive, because, in his words, "the influence of something is really difficult to evaluate" and "you have to precisely define the boundary of unwanted capability" [7]. The exposure a company carries is therefore set by an architecture decision taken long before any employee misbehaves.
Camilla Hrdy, a trade secret scholar at Rutgers, offers a different read. She told Business Insider these cases do not immediately call for novel legal solutions, since the familiar remedies still apply: orders to stop using the secrets, orders not to disclose them, orders to protect them, plus damages measured as actual loss or, in some cases, royalties [5]. She is right on the doctrine, but the enforcement problem she skips is real: an order to stop using a secret is cheap to write and expensive to verify once the thing to be stopped lives inside model weights whose boundary someone must first define [7]. Hrdy concedes the part that matters here: employees plugging what they know into AI "could be a real loss of control," and that is new [4].
What the record does not contain is the fact pattern that would settle which branch applies. Apple did not say whether the former employee ran a one-off AI-assisted simulation or did something that could affect a broader model [8], and an Apple spokesperson did not return a request for comment [9]. The nearest prior case is xAI's suit against OpenAI, which alleged that former engineer Xuechen Li kept xAI's entire codebase in personal cloud storage and connected that storage to his personal ChatGPT account as a "Source," giving OpenAI a means to access the files [10]. A judge dismissed it in June [11]. So the supplied record holds no judicial test of the irreversibility theory at all [12].
That gap sets the sequencing. This quarter's decision turns on inventory rather than on unlearning, which remains a research problem: can you enumerate which agents your staff have connected to which repositories? That inventory is what determines whether a future remedy is a delete command or a discovery fight. Liu's own interim suggestion is a detection system that flags sensitive requests or sensitive information passing between agents and triggers a hard stop, which he acknowledges is not true unlearning but is more practical [13]. The cost of that control is friction imposed on exactly the employees using agents most productively, and the cost of skipping it is that the company's answer to "where did the secret go" has to be reconstructed under a deadline set by someone else's brief.