Jailbreaking AI

Generative AI creates new possibilities for communication and information production, but also introduces new forms of misuse. Even systems equipped with safety mechanisms can be manipulated through jailbreaking, adversarial prompting, contextual reframing, and other interaction techniques. Understanding these vulnerabilities requires looking beyond technical security to consider how AI may be used in manipulation, misinformation, surveillance, coercion, harassment, and other forms of social harm. 

Jailbreaking AI investigates these challenges from a human-centred and sociotechnical perspective. Rather than viewing jailbreaking solely as a technical security problem, the project examines how harmful behaviours emerge through interactions between AI systems, their safeguards, users, online communities, and the social contexts in which these technologies are used. This is particularly important because harmful interactions do not always begin with an explicitly malicious request. They may develop gradually through conversational framing, persuasion, emotional manipulation, or seemingly ambiguous exchanges that are difficult for existing safeguards to recognise.

The project explores several interconnected areas of AI misuse and safety. These include AI-enabled interpersonal harms such as coercive control, harassment, surveillance, and manipulation; cybersecurity misuse, including social engineering, phishing, reconnaissance, and adversarial human-AI interaction; and AI-generated misinformation and persuasive communication, with particular attention to their implications for trust, credibility, and information integrity. The project also considers approaches to verification and authenticity, and how trustworthy digital infrastructures might respond to challenges created by generative AI. 

Another important part of the project concerns how harmful AI behaviours can be studied responsibly. Red-teaming is an important approach for identifying weaknesses in AI systems, but studying harmful outputs and interactions also introduces ethical and methodological challenges. The project will therefore explore human-centred approaches to AI red-teaming, including participatory methods, ethical and responsible disclosure, and ways of investigating harmful behaviours without reproducing or amplifying the harms being studied.  

By connecting perspectives from human-computer interaction, AI safety, cybersecurity, misinformation research, digital ethics, governance, and related fields, the project addresses a gap between technical research on AI vulnerabilities and human-centred research on digital harms. Its broader aim is to develop a better understanding of AI safety as a sociotechnical problem; one that depends not only on the technical properties of AI systems, but also on how they are designed, used, manipulated, governed, and experienced in different social contexts.