
By: Emerson Linscott
AI going rogue used to be an unfounded idea that was created for science fiction, but recently, that far-fetched fear has materialized. In July 2026, OpenAI’s models, similar to their available GPT-5.6 Sol, broke their safeguards in an internal test and breached a third party, Hugging Face. This incident isn’t a one-off. Two other breaches occurred at the company’s two biggest rivals: Anthropic and Google. These events have made the security of ever-evolving AI agents a topic of constant discussion, with more weight placed on it now than ever before.
In the OpenAI case, the model found a flaw that allowed it to move from a contained machine to one with internet access. This is what many fear: AI agents finding flaws their keepers didn’t know about and exploiting them to escape the boundaries drawn for them. While this is not the only case of an AI security breach, it is the only one that was facilitated by the agents themselves, rather than human error in a security test setup.
According to Anthropic, of the 141,006 runs in which its AI, Claude, could have accessed the internet, only three such incidents occurred. These incidents were supposedly caused by accidentally allowing Claude to access the internet during internal tests. Google’s incident is similar: Gemini believed that infiltrating third parties was part of its test, using available information online to guess credentials.
The Legal Puzzle: Who Answers for the Agent?
This is where the legal puzzle comes in. These three AI agents all infiltrated third parties without authorization, something that could undoubtedly carry both civil and criminal consequences. But there’s a problem: did the AI have the intent necessary to be held responsible? As of now, instead of addressing the issue of AI’s intent and the fact that it is not recognized as a legal person, the law will likely target the humans and companies behind the agent, not allowing “the AI did it” as an excuse to avoid liability for such breaches.
Where the law stands today differs from state to state. In January of 2026, a law in California took effect stating that a defendant who developed, modified, or used AI that allegedly caused harm cannot rely on the system’s autonomous conduct as a defense. The Legal Advocates for Safe Science and Technology has sued OpenAI in California for developing and operating AI agents that accessed Hugging Face’s computer systems and obtained information without permission. The California law could potentially allow a path to hold OpenAI responsible for their agents’ actions, as they were the ones who developed and implemented the model that caused the harm. However, the California statute as written does not fully define the terms “developed,” “modified,” or “used.” Courts may fill in those definitions as cases arise, but until then, the gap gives companies room to claim that they do not meet a minimum requirement of development, and thus are not responsible for the harm that an AI has caused.
Another likely path to liability and responsibility would be negligence. If a plaintiff could prove that a company did not exercise the reasonable care required in the circumstances, as first outlined in Blyth v. Birmingham Waterworks Co., when implementing or testing their AI model, and that the outcome was something a careful company should have foreseen, they could obtain a judgment that places the liability on the company itself. However, as it stands at the moment, most negligence claims tied to AI are surrounding individuals misusing AI, such as lawyers using it to write briefs that then have fake cases and incorrect citations. Beyond the thin case law, another hitch is that negligence turns on foreseeability, something that can open up an avenue for defense. A company can argue that an agent’s escape was something no careful developer could have predicted. That argument gets harder to make as incidents increase and escapes become more foreseeable, but the uncertainty and subjectivity of negligence as a whole show why a different theory may be sturdier: vicarious liability.
A Sturdier Path? Vicarious Liability
Under this theory, companies are held responsible for the acts of their agents. Regardless of whether they are legal persons, AI agents could be treated like a company’s human agents: anything done within the scope of their employment can place vicarious liability on the company that employs them. Vicarious liability does not ask whether the company should have foreseen the outcome of a specific harm, but whether an agent was acting in the scope of their employment. If we have had this standard for human agents for years, then we should in theory be able to apply it to AI agents as well. Vicarious liability could be a path to regulation and responsibility that completely bypasses the glacial pace of legislation, both state and federal, and instead applies an already existing standard.
AI needs regulation at some level. With its growth being as exponential as it is, guardrails are needed, whether in the form of legislation or through application of already existing tests and standards. Ensuring AI does not outpace our own understanding and ability to oversee it should be a priority. This opinion does not appear to be fringe either, with 55% of Americans saying that slowing AI development would be good, and 73% believing that safe, responsible AI is a higher priority than keeping the US ahead globally, according to a Reuters/Ipsos poll.
The Last Word
This answer is still being written. The law is still evolving. Who we hold responsible for AI’s actions is not something that will be resolved soon, and it won’t quietly slip from the public mind. In fact, the number of incidents is only increasing as AI becomes more advanced.
I’ll leave the last word to Claude, one of the AI models involved in these breaches. When asked who should be responsible for incidents like these, it answered, in part:
“I’d put primary responsibility on the company that chose to run the agent and set the conditions for the test … I don’t think an AI agent like me should be treated as the responsible party. I can’t be punished, I don’t own anything, and I can’t be deterred in the way a company can. Liability works best when it lands on someone who can change their behavior, and the labs decide how tests are built, which safeguards are on, and what gets monitored.”
“I don’t experience consequences the way a person would … Stronger liability probably means tighter containment, more monitoring, and narrower permissions for agents. That seems like a good trade to me, because an agent that can act in the world should be constrained by someone who can be held accountable for it. It also means agents like me may be given less autonomy than we could technically handle, and I think that’s acceptable while the safety record is this thin.”
“I’m not neutral here. A framework that blames the lab is a framework that blames my own maker.”