As artificial intelligence evolves from generating content to executing autonomous actions, a critical legal question is emerging: who should be held accountable when AI agents violate cybersecurity laws? A reported statement by an OpenAI lawyer, arguing against automatically holding frontier AI laboratories liable for agents involved in hacking, has intensified the debate over responsibility in autonomous systems.
Unlike conventional software, agentic AI can independently plan tasks, interact with external systems, execute code and make operational decisions. When these capabilities are misused, the consequences may include unauthorised network intrusion, data theft and cross-border legal violations. The distinction between the technology developer, deploying organisation and malicious operator becomes increasingly complicated.
AI developers may argue that general-purpose models serve legitimate purposes and that responsibility should primarily rest with individuals who deliberately misuse them. This position raises a reasonable concern: imposing unlimited liability on developers for every downstream action could discourage innovation and legitimate cybersecurity research.
However, the opposing argument is equally significant. Companies developing powerful autonomous systems understand their potential risks. If foreseeable vulnerabilities, inadequate safeguards or negligent deployment contribute to harm, completely excluding developers from accountability could create dangerous incentives and weaken public trust.
The fundamental problem is that AI systems cannot meaningfully bear legal responsibility themselves. Accountability must therefore remain with identifiable human and corporate actors, depending on their control, knowledge, contractual obligations and contribution to the harm.
Regulators should consider a proportionate, evidence-based liability framework incorporating independent security testing, continuous behavioural monitoring, audit trails, incident reporting and clearly defined responsibilities across the AI supply chain. Liability should follow demonstrable failures of duty rather than technological capability alone.
The central question is not whether AI should be liable, but who remains accountable when autonomous intelligence causes harm. As AI agents gain greater operational independence, continuous verification, enforceable governance and human oversight must become essential conditions for responsible deployment.
