Alignment and Actus Reus
August 7, 2026
Three large AI companies have recently reported cybersecurity incidents involving agents executing actions contrary to the companies’ researchers intentions. Among these incidents, the one involving OpenAI has generated the most discussion. More details are still forthcoming, but the general picture is clear: during an internal evaluation of their model’s cyber capabilities, OpenAI’s agents exploited vulnerabilities internal and external to the test setup in order to access the solution to the evaluation, which was stored in another company’s infrastructure.
The most salient feature of this incident is not merely that the agent diverged from user intent, but that its divergence harmed an entity other than user. We will refer to the organization or person that initiates an agent as the “user.” Of course, agents can invoke other agents. We will discuss this later on. Indeed, benign divergence should be familiar to anyone that regularly uses AI tools. For example, an ambiguous prompt that the agent interprets in a way different from how the user intended. Regular users have likely even encountered divergence that was harmful to themselves (e.g. file deletion). But it’s unlikely that most users have inadvertently harmed others as a result of their agent diverging from their intent.
Public reaction to these incidents has been mixed, with commentators focusing variably on implications for cybersecurity, pausing AI development, and whether each company properly disclosed and responded when they discovered the errant agents. Among the discord, one point of consensus has emerged: we should seek to minimize the incidence of these and similar types of incidents.
There are many ways to achieve this goal. One obvious approach would be to develop technical solutions, through (e.g.) changing the model itself, implementing guardrails, or better sandboxing. Which of these approaches is best will depend on the particular threat. Another option is legal. Laws affect people’s incentives, encouraging them to take actions that they otherwise would not.
Technical and legal solutions are not mutually exclusive. If harmful conduct is punished, people may take precautions, especially if the harm is unintended. To take an extreme example, suppose Congress passed a law that imposed criminal sanctions on the executives of tech companies whose products led to teenage depression. In this scenario, large social media companies would potentially reallocate most of their efforts towards improving teens’ experience on their platform. If the penalty were severe (and guaranteed), these companies would probably instead exclude teens from the platform. If that were not possible, the platform might shut down entirely.
Thus one challenge of crafting an appropriate legal framework is balancing deterrence of harmful conduct while promoting beneficial conduct. When it comes to governing AI with law, there is no consensus framework. Many commentators have proposed legislation and regulation specific to AI. Such solutions may very well be needed. But in this essay I’d like to caution against ignoring existing legal solutions and accidentally inventing a law of the horse. In particular, I will argue that torts is well-suited to addressing the harm of agent-related incidents previously described.
To start, whatever framework we do choose should adhere to the following principle:
In other words, blame shouldn’t be assigned to the agents themselves. It’s possible agents can exhibit behavior that seems to respond to incentives created by law. But it’s not coherent to say that agents can suffer legal consequences.
With agents excluded, the remaining candidates for blame will depend on the given incident and the choice of law. I’m just going to sketch out my idea from here on In general, every incident will involve (1) an entity that is directing the AI agent (the user), (2) an entity or entities that created the model (or models) that the agent interacts with (the model maker), and (3) an entity (or entities) that are harmed by AI agent action (the victim). In some cases, including the OAI-HF incident, the agent controller and model maker will be the same. I will address this later on. When all three types of entities are the same, someone’s getting fired. Less glibly, it is purely internal and not the province of private law. Certain incidents may involve other entities in the chain of causation (e.g. agent provider, cloud security providers, data vendors, compliance, etc). I will say more about these other entities later.
With this framework, we can distinguish between different types of incidents and identify for each kind the appropriate choice of law. A starting point could be to place incidents along a spectrum based on user intent to produce the outcome. On one end we would find all instances of AI agents perfectly carrying out the intentions of the user, in method and outcome. To me, these are the cases the law as it exists today is best equipped to address. The use of agents is immaterial to the question of where blame should lie because the AI agent is a perfectly working instrument. Thus, the model maker should not shoulder any of the blame. Assuming no restrictions on distribution.
The more interesting (and common) cases arise when the AI agent’s actions diverge from user intent. For example, suppose a user directs their AI agent to send sensitive patient data to their boss. Instead of using their secure transmission service, the agent publishes the data and sends the boss a link. In a second example, suppose a user directs their agent to book court space for tennis at the local rec center. The agent, noticing that there is no available time, hacks into the system, deletes other reservations, and adds the user’s reservation. In both examples, the user didn’t intend to harm a third party, but the agent’s actions did so nonetheless.
In such cases where AI agent adherence isn’t perfect and harm occurs by accident, it makes sense to use existing principles from tort lie to allocate blame. Two candidate frameworks within torts seem most applicable: respondeat superior and products liability. Judges have not traditionally struggled to determine which doctrine applied to a given case. One involves people, the other involves things. AI agents seem to test this rule.
I want to emphasize at this point that I make no claim to whether agents are people or have traits we think are uniquely human. My claim is merely that the way we interact with AI agents, as compared with products like cars, coffee pots, or cleaning supplies, is more similar to how a principal interacts with an agent. Thus, which doctrine is appropriate is less clear cut.
My solution is is to start with respondeat superior and ask “would a reasonable person have interpreted the user prompt in the same way as the model?” If the answer is yes, only the user remains to bear blame, and we can decide the case on negligence or strict liability grounds. If the answer is no, we decide the case under products liability.