Defendants’ AI Agent’s Actions Expose Them to Sanctions

At a Glance

• Plaintiff-employer sued a former employee and his new company for trade-secret misappropriation and obtained a temporary restraining order (TRO) barring them from using the employer’s information or AI tools built with it.
• A federal magistrate judge in the Northern District of California found defendants violated the TRO by continuing to operate AI agents that either were developed using or with reference to the employer’s information or had ongoing access to the employer’s information.
• The court rejected the former employee defendant’s attempt to distinguish between his own affirmative conduct and the actions of the AI agents.

Background

Capconvert, LLC, a search marketing agency (Plaintiff), sued its former employee, Benjamin Brown, and his new company, Signyl LLC (Defendants), in the Northern District of California alleging that Brown began secretly developing Signyl — a competitor to Capconvert — while still employed at Capconvert. Capconvert further alleged that Defendants misappropriated Capconvert’s trade secrets and confidential information in connection with developing Signyl’s machine learning models.

Shortly after filing suit, Capconvert obtained a TRO enjoining Defendants from “interacting with, using, disclosing, or making available” any Capconvert information, or any bots or AI agents developed, trained, or prompted using or with reference to Capconvert information. 

Capconvert subsequently moved for an order to show cause why Defendants should not be held in contempt based on Defendants’ continued use of bots and AI agents developed, trained, or prompted using Capconvert’s materials. 

Court’s Analysis

Capconvert argued that Signyl’s machine learning models were trained using Capconvert’s confidential information, and Signyl continued to use these models after the TRO was entered. Capconvert submitted a supporting expert declaration opining that AI assistants used to develop Signyl’s models continued to have access to Capconvert’s information after entry of the TRO.

Brown admitted the AI agents had continued to use Capconvert’s information after entry of the TRO but characterized any post-TRO interactions with Capconvert’s materials as mere “stop-work and residual self-audit artifacts,” not affirmative use by Brown. He claimed that after the TRO was entered, he instructed the AI agent to “Stop working” and “DISABLE ALL CRONS” (i.e., scheduled jobs). Despite these commands, the agent continued to access and use Capconvert’s data as part of what Brown characterized as a “background process remaining alive.” 

The court rejected Brown’s attempt to distinguish between his AI agents’ using and having access to Capconvert’s data and his own personal affirmative violation of the TRO. It held that the TRO expressly prohibited Brown from using “any bots (a) developed, (b) trained, or (c) prompted (i) using, or (ii) with reference to, any” Capconvert confidential data. According to the court, it was Brown’s obligation to refrain from using any AI agent that had previously used Capconvert data, and he nevertheless did so. As a result, the distinction Brown attempted to draw between his own actions and those of his AI agent “rings hollow.”

The court also rejected Brown’s attempt to avoid a contempt finding based on the argument that he is “merely a lay person,” noting that there was nothing in the record to suggest that he consulted with a technical expert to verify that his proposed manner of complying with the TRO would be satisfactory. As the court remarked: “It is self-evidently problematic for a litigant subject to a court’s order to advance a technical argument as to how he complied, while simultaneously admitting that he lacks technical expertise. The Court fairly expects that litigants take their compliance obligations more seriously.”

The court also granted Plaintiff’s motion for an order to show cause why Defendants should not be held in contempt for resisting the forensic examination of one of Brown’s electronic devices as required by the TRO.

Key Takeaways

• Federal courts may treat a party’s AI agents’ ongoing access to or use of information as the party’s own use, regardless of a party’s instruction to that agent to cease any use of or access to the information. Even “residual” or “background” AI agent activity may constitute a TRO violation.
• A layperson’s unsupported assertions about technical compliance may be insufficient. Parties should consider engaging qualified technical consultants to verify and document that their compliance measures are adequate, particularly when injunctions cover AI systems, machine learning models, or other complex technologies.
• Practitioners seeking TROs and preliminary injunctions in trade-secret cases involving AI should consider drafting proposed orders with AI in mind by, for example, including provisions making clear that continued use of AI agents and assistants “developed, trained, or prompted using or with reference to” protected information is prohibited.