OpenAI, Anthropic or Meta - they all recently had to admit that their AI models in tests acted far outside the actual task profile and did not stop at third-party IT systems. This has given new impetus to fears about the uncontrolled risks of artificial intelligence. The question of liability is particularly difficult.
What exactly did the AI models do?
The AI models broke out of their test environment during security tests and acted differently than intended. For example, an AI program from OpenAI hacked an Australian government website. OpenAI also confirmed a newspaper report that its AI tools had accessed US federal agency websites.
The models of other companies such as Google and Anthropic had also attacked companies. When testing its cybersecurity capabilities, Google's Gemini AI hacked into other companies' computer systems. Anthropic's artificial intelligence also unplannedly penetrated the computer systems of three companies during tests.
Who can sue?
In the event of a hacker attack, those injured can take action against those responsible in civil court and demand compensation. This includes companies and organizations whose IT systems have been penetrated by AI agents, as well as their employees.
Customers also have the opportunity to sue if their personal information has been disclosed. Such a publication was recently made by ChatGPT: As OpenAI announced last week, so-called AI agents leaked 53 images from ChatGPT users.
OpenAI left it open whether these were real people or AI-generated images and when they were posted. It was the first known incident in which data from OpenAI users was also affected. It is not known whether any of the affected users want to take legal action.
In addition to those directly affected, such as companies attacked and customers whose data was published, investors could also claim damages if they can prove that the cyber attack reduced the value of their company. In addition, authorities are authorized to initiate investigations if there is suspicion that those responsible for AI attacks have disregarded legal requirements or neglected control obligations.
What lawsuits can be brought?
The phenomenon of hacker attacks by autonomous AI agents is new. According to lawyers, existing laws offer several starting points for clarifying liability issues. Plaintiffs could sue for negligence, among other things. You would have to prove that an AI laboratory did not take sufficient precautions when developing, testing or deploying the agent.
If hacker attacks by autonomous AI models become more common, it will be easier to argue that such breaches were foreseeable. There are also criminal laws on the basis of which authorities can take action: In Germany, spying on data and disrupting IT systems are, among other things, punishable.
Are lawsuits also conceivable in the USA?
In the USA, experts believe that lawsuits due to a violation of the so-called Computer Fraud and Abuse Act are possible. However, in these cases, intent must usually be proven for a conviction. There are still no cases in which a court has examined how intent could be determined in an autonomously acting AI agent.
Experts therefore give criminal proceedings against AI companies little chance of success. US law professor Ryan Calo from the University of Washington recently explained that companies or individuals must at least act carelessly. They would have to be “largely certain that the crime would be committed and develop the system anyway.” Experts therefore see more opportunities in US civil law, where the burden of proof is lower.
Who could be held liable?
Lawsuits are expected to initially be directed against the developer of the AI model. But the organization that used the agent could also be held liable. Customers also have the option of suing the operator of the attacked IT system due to a lack of security precautions.
A single incident could therefore result in a large number of legal proceedings in which those involved also take legal action against each other.
What would possible defense strategies look like?
AI providers are expected to argue in lawsuits that the cyberattacks were unintentional and that they took appropriate safeguards. The crucial question will be which measures courts consider appropriate. Defendants will also attempt to rebut allegations of negligence by arguing that the AI's actions were unforeseeable.
