Apple raises new concerns around clawing back trade secrets from an AI

Here’s a novel problem companies may have to deal with in the AI era.

A nefarious Big Tech employee leaves for a competitor with trade secrets, feeds them to an AI agent or model while employed by the competitor, and runs some tests using those secrets.

Maybe the bad-apple employee then creates a new solution using that confidential knowledge, which the competitor benefits from.

Or perhaps those secrets are stored in some knowledge base that an AI could retrieve if the nefarious employees’ colleagues have a relevant question.

Apple raised that possibility in a supplemental brief filed Monday in support of its request for expedited discovery in its trade-secret lawsuit against OpenAI.

In the filing, Apple’s attorneys said a former employee’s use of company secrets while employed by OpenAI and his “use of AI agents to learn to run simulations raise concerns extending beyond ordinary document theft.” “Where trade secret information is fed into an AI agent or model that ‘learns’ from it, such ‘learning’ may create irreversible and continually propagating uses of the trade secret — harm that, at a minimum, is uniquely challenging to undo and requires prompt investigation,” Apple’s lawyers wrote.

The continued use of confidential information by a rival company is not a new problem.

Artificial intelligence, however, is introducing a new wrinkle to the matter: How should companies regain control of their secrets after they’ve entered an AI system at a competing organization? “Employees are already real loose cannons, walking around with knowledge in their heads,” Camilla Hrdy, a law professor at Rutgers whose work examines trade-secret law and generative AI, told Business Insider. “Now they’re taking that knowledge and plugging it into AI, and that could be a real loss of control.

That is new.” Elon Musk’s xAI raised a related but distinct AI-linked concern when it sued OpenAI, accusing Sam Altman’s company of poaching staff to steal Grok’s underlying technology.

That lawsuit said that, while Xuechen Li, a former xAI engineer, “had xAI’s entire codebase stored in his personal cloud storage account, Li also had his personal ChatGPT account directly connected to his personal cloud storage account, set up as a connected ‘Source’ in OpenAI’s ChatGPT.” It added: “OpenAI had a means to access Li’s files, which included the stolen copy of xAI’s entire source code, through its ChatGPT service.” A judge dismissed the lawsuit in June.

Hrdy said these cases don’t immediately call for novel legal solutions.

Potential remedies often include telling a company to stop using the trade secrets, not to disclose any secrets, and to take steps to protect said secrets.

There are also damages to be assessed, Hrdy said: actual losses incurred from losing those secrets, or, in some cases, royalties to be paid to the affected company.

Stopping trade-secret use by a rival company could pose technical challenges, depending on exactly how an employee applied confidential information to an AI system.

Sijia Liu, a computer science professor at Michigan State University, co-authored a paper on “machine unlearning” — the process of removing the influence of certain data, or capability, from an AI model.

He told Business Insider that if a document containing sensitive information is stored in a repository an AI system retrieves from, the remedy could be as relatively straightforward as deleting the file.

On the other hand, if sensitive information were used to train or fine-tune a model, it would require an entirely different, and likely resource-intensive, process. “The second case could be more difficult because the influence of something is really difficult to evaluate,” Liu said, adding that “you have to precisely define the boundary of unwanted capability.” A more immediate approach to containing secrets could be to build a “detection system” that flags sensitive user requests or sensitive information being passed between agents, Liu said.

The detector could then trigger a hard stop in response to the request.

Liu said that’s not “true unlearning,” but it is more practical.

To be clear, Apple did not say how the former employee may have used trade secrets with an AI, whether it was a one-off AI-assisted simulation or whether there was training that could affect a broader model.

An Apple spokesperson did not return a request for comment on this story.

Either way, AI may be bringing up new ways for companies to lose control of their secrets.