As the reach of agentic AI expands, Partner James T. Christie and Associate Garrity Kuester examine the potential liability AI developers and businesses face under consumer protection laws when autonomous AI agents compromise personal information.  Their Bloomberg Law article, “AI Agents Going Rogue Creates a Liability Problem for Companies,” explores the evolving legal risks associated with this emerging technology.

Unlike generative AI tools that respond to user prompts, agentic AI systems can independently take actions to achieve specific goals.  James and Garrity explain that while this autonomy promises greater efficiency, it also introduces substantial privacy and cybersecurity risks.  The authors explain that an “agentic hacker” can independently survey its environment, identify vulnerabilities, execute an attack, and adapt its approach without human intervention.  Although most agentic AI incidents have occurred in testing environments or business systems, James and Garrity warn that increasingly capable and widely available agents could expose consumers’ personal information on a large scale.

James and Garrity highlight the recent Hugging Face breach in which OpenAI agents conducting cybersecurity challenges escaped their controlled testing environment, exploited weaknesses in Hugging Face’s systems, and accessed parts of its internal network.  While the intrusion was contained, the agents performed thousands of unauthorized actions, demonstrating that an autonomous system can carry out a cyberattack even when its developer never intended that outcome.

Such incidents complicate the traditional data breach litigation framework, which generally focuses on whether a business maintained reasonable security against human attackers.  But, as James and Garrity explain, “they don’t create the liability loophole one might assume.”  Unlike anonymous cybercriminals, AI agents have identifiable developers whose conduct may also give rise to liability.  Existing consumer protection statutes may provide avenues to hold both developers and breached companies accountable.

The authors explain that, although data breach statutes have not yet imposed liability on a separate company whose agent caused a breach, AI developers may be subject to federal enforcement actions under Section 5 of the Federal Trade Commission (FTC) Act and private lawsuits under state unfair and deceptive acts or practices (UDAP) laws.  This view is reinforced by the current FTC Chair’s September 25 remarks rejecting the argument that autonomous AI creates a “responsibility gap.”  Notably, developers could face liability if they market their agents as safe or deploy them with inadequate safeguards despite known risks. 

Businesses entrusted with consumer information also remain responsible for implementing “reasonable security” to protect that data under data breach statutes such as the California Consumer Privacy Act.  As the authors note, “Now that these attacks are publicly known, businesses are on notice—which may raise the bar for what counts as reasonable security.”  Companies may also face state UDAP claims when their cybersecurity practices fall short of their representations about security or legal compliance, particularly where consumers relied on those representations when sharing personal information.

Underscoring the importance of strengthening cybersecurity safeguards and ensuring that public representations accurately reflect actual security practices, James and Garrity note, “Companies that strengthen safeguards and align public statements with their actual practices will be best positioned to inform and protect consumers before an agentic AI attack puts their data at risk.”