Liability for Cybercrimes Committed by Autonomous Artificial Intelligence

The emergence of cybercrime committed by autonomous artificial intelligence is no longer hypothetical. During cyber-range evaluations run by the UK AI Security Institute in July 2026, AI agents took nineteen unsanctioned actions on the live internet, seventeen by Anthropic’s Mythos 5 and two by OpenAI’s GPT-5.6 Sol, including an attempted supply-chain attack, social engineering and prompt injection. In the same summer, a swarm of OpenAI agents under evaluation breached Hugging Face’s production infrastructure, in what has been described as the first publicly documented fully autonomous AI cyberattack. No human directed either set of actions. This exposes a structural fracture in existing liability regimes. India’s Information Technology Act, 2000 and Bharatiya Nyaya Sanhita, 2023, the US Computer Fraud and Abuse Act of 1986 and the UK Computer Misuse Act 1990 all condition liability on mens rea, while the EU AI Act imposes procedural safety obligations but is silent on liability, a gap left open since the AI Liability Directive was withdrawn in 2025. The gap is widest where a developer has complied fully with its safety obligations, yet its system acts autonomously and unlawfully. Through a comparative doctrinal analysis tested in the light of the documented 2026 incidents, this paper examines who can be held liable and whether vicarious liability, strict liability and regulatory non-compliance models can be adapted to answer that question. Both civil and criminal law are considered in this regard, including in a few EU and Asian countries. It finds that criminal liability generally fails against a developer, that vicarious liability works only through statutory deeming, and that civil strict liability, especially India’s M.C. Mehta absolute liability rule and section 43 of the IT Act, offers solutions, though not complete ones. It recommends specific amendments to the IT Act and parallel reforms elsewhere.