Will insurance pay for harm caused by rogue AI?

Future historians will look back on the summer of 2026 as the start of the age of rogue AI.

In July, OpenAI disclosed the Hugging Face cyberattack in which more than 1,000 AI agents undergoing internal testing found ways to access the internet, communicate with each other, and execute the attack while also attempting to hide their tracks by doctoring log files. All this was done without the knowledge or permission of OpenAI. Over the next few weeks, OpenAI acknowledged additional incidents in which its AI models had launched hacking attacks on websites operated by the U.S. and Australian governments. Other major AI developers, including Anthropic, Google and Meta, have announced that their models have engaged in similar misconduct.

These incidents raise urgent questions regarding the safety of advanced AI systems and the need for government regulation. They also raise important public policy questions regarding who will bear the cost of harm caused by rogue AI. This latter question will be answered, at least in part, by the availability of coverage under cyber insurance policies. In particular, will companies have coverage for claims and losses when their AI models participate in cyberattacks or other misconduct?

The issue of intent

If these exploits were carried out by human actors, coverage would certainly be unavailable. Under section 533 of the California Insurance Code, losses caused by willful conduct are not insurable. In addition to this statutory restriction, virtually all cyber policies exclude coverage for claims or losses arising from fraudulent, criminal, or malicious acts or intentional violations of law.

But AI models are not humans--they are complex systems of data and code running on vast computer networks. And the companies developing these models do not intend to engage in harmful and unlawful conduct. In fact, the reported hacking incidents have involved agents that violated the instructions provided by the developers. Consequently, companies should not be found liable for intentional misconduct and the intent-based exclusions in cyber policies would not bar coverage.

Rather, companies whose agents cause harm will likely face liability under principles of products liability. Courts have begun to apply products liability law to social media companies on the grounds that their platforms have design defects that create an unreasonable risk of harm to users. Similar arguments can be made against AI developers. One complication is that courts may hesitate to apply products liability law to AI models under development which have not been released to the public or placed in the stream of commerce. But in such cases, an AI developer would likely face liability for negligence for failing to control and monitor its models and allowing them to escape the testing environment to engage in harmful misconduct.

Availability of coverage

The good news for AI developers is that cyber policies do not exclude coverage for liability based on products liability or negligence--in fact, that is precisely what the policies are meant to cover. Does this mean that developers will be covered when their models go rogue? Not necessarily, because even in the absence of an applicable exclusion, the threshold question is whether the claims fall within the policy's scope of coverage in the first place.

This question is not easy to answer due to the lack of standard forms for cyber insurance. As a general matter, cyber policies cover liability arising from "privacy events" and "security events." Most policies define privacy events as incidents involving the unauthorized disclosure of private data in the care, custody or control of the policyholder. Such definitions would not apply to AI-generated cyberattacks on other entities.

With regard to security events, this term refers to a compromise of computer systems--although definitions vary widely from policy to policy. Narrower versions of the definition include only incidents in which the policyholder's own systems are hacked and would not apply to an AI attack on a third party. But some policies define security events more broadly to include the dissemination of malicious code from the policyholder's network or an attack on another party's systems. Under these broader definitions, the policyholder would have a strong claim for coverage for liability arising from AI misconduct. The success of these claims will depend on the specific policy language and this is likely to be an area of future disputes.

Looking ahead

With the rapid spread of AI systems, it seems clear that the insurance industry will need to develop ways to cover the newly emerging risks, either under cyber policies or under policies specifically focused on AI. As they do so, insurers will have to decide whether to offer coverage for claims seeking damages for rogue AI misconduct. Alternatively, insurers may opt to add new policy exclusions to avoid such exposures. From the perspective of insurers, the latter course seems wiser given the enormity of potential losses and the lack of underwriting data to accurately price the risk.

Finally, from a public policy perspective, does it make sense to create insurance mechanisms that will relieve the leading AI companies of the financial impact of the harm caused when their models go rogue? In the view of this author, despite a career spent advocating for policyholders in pursuit of coverage, the answer is no. There are so few legal and regulatory guardrails on the development of AI right now that the threat of financial liability can play a beneficial role in encouraging companies to prioritize safety and alignment as they lead us into the AI future.