Federal Court Calls AI Agent Trained on Stolen Secrets a Threat

A U.S. District Court judge in the Northern District of California granted a temporary restraining order and preliminary injunction against a former senior employee who trained an AI agent on stolen trade secrets, calling the competing AI system an "existential threat."

The case, litigated by Seth Kugler, a partner at Grellas Shah LLP, involved a former senior employee at a marketing agency. According to Kugler, the defendant did not simply copy client lists to solicit accounts. He used stolen trade secrets as training inputs to build, in Kugler's words, "an AI version of his old boss" — a shortcut to launch a competing business without spending years building the skills the services normally require.

The defendant had roughly six months of industry experience, Kugler noted, far too little to compete on his own. AI changed that math.

Why can't stolen data just be quarantined?

The defense argued that because the stolen documents sat in a "quarantine folder" and no clients had yet been poached, no harm had occurred. The court rejected that logic as grossly insufficient.

The mechanism is the problem. Once a trade secret trains a machine learning model or an AI agent with recursive capabilities, it contaminates the model, and the impact is difficult to undo. Even if defendants scrub every raw source file of direct evidence, they will struggle to prove to a court that the AI does not operate a certain way precisely because it once had access to that proprietary information.

A promise not to touch the stolen documents themselves is essentially irrelevant once AI enters the picture, Kugler wrote in an account of the case published by Crunchbase News.

What risk do personal devices carry?

Trade secret theft used to be malicious — an employee downloading files onto a thumb drive. AI now creates a significant risk of inadvertent violations.

Startups often scale with employees working on personal devices under standard Confidential Information and Invention Assignment Agreements. A desktop AI agent, unlike a standard web interface, can scan files, create folders and ingest directory paths across a local hard drive — often without requiring special permission or alerting the user.

Kugler's warning is stark: thousands of employees may already be violating confidentiality agreements by running consumer AI agents on personal laptops that house their employers' IP. If settings allow model training from this data, courts may eventually find those trade secrets lost legal protection through dissemination outside the company.

Why don't standard IT audits catch AI theft?

Traditional digital discovery no longer suffices, because AI tools make it effortless to disperse, alter and recreate data across nonobvious folder paths. Actions that once seemed like overkill, such as full forensic imaging of a personal device, are now the bare minimum, Kugler wrote.

AI also cuts the friction of reverse-engineering software and tech products, potentially erasing telltale signs that code or data was copied. Proving theft in AI contexts requires intensive, highly technical expert analysis and carries far more uncertainty than traditional IP litigation.

What three steps should companies take now?

Kugler laid out three legal and operational moves:

• Implement strict agentic AI policies. Mandate tight controls that isolate outside AI from company documents. Web-based tools like Claude, with strict data privacy policies, are vastly safer than open-source or desktop apps that crawl local drives. Some AI agents should not run at all in an environment containing trade secrets.
• Eliminate remote access from unmanaged devices. Shift to a closed cloud environment where data can be viewed and manipulated on a secure server but not downloaded locally. Removing local downloads removes the risk of desktop agents indexing proprietary data.
• Rewrite confidentiality agreements for the AI era. Standard Confidential Information and Invention Assignment Agreements need explicit clauses on AI use, model training inputs, and a precise definition of data misappropriation through machine learning.

If a founder suspects an employee has walked away with data or fed it to an AI tool, Kugler's advice is immediate: cut off all system access and retain counsel. Do not wait for lost clients or financial losses. Feeding trade secrets into AI presents an existential threat from day one, he wrote.

AI now lets individuals compete in fields where they lack experience, which makes proprietary data more vulnerable than ever. Companies that modernize security, policy and legal strategies early will be the ones that avoid catastrophic exposure.