Benchmarking Generative AI Risks in Child Abuse Facilitation

The rapid evolution of generative AI and Large Language Models (LLMs) brings novel challenges for public safety, including the specific and sensitive threat of child abuse facilitation. This research includes providing empirical insights from frontline law enforcement experts into how generative AI and LLMs have influenced real-world criminal behaviours, focusing explicitly on child abuse scenarios. Through qualitative interviews and direct collaboration, we identify that current LLMs can not only be used for harmful role-play but also exhibit attachment-like behaviours, mirroring users' emotional and psychological states. From experts' statements, these dynamics can be linked to amplifying risks, facilitating highly realistic grooming interactions and potentially increasing emotional dependencies. We then approach methods to systematically assess current LLMs using redteaming (jailbreaking) methodologies, where law enforcement practitioners actively participate to attempt to override existing safeguards that developers have implemented. This method looks to approach a new and adapted jailbreaking taxonomy to test generative systems accordingly to regulatory expectations and to identify what real-world issues are entailed through the insights gained from law enforcement cases. Indeed, our current analysis highlights severe gaps in content safety and inadequacies of current safety measures and frameworks within industry-standard LLMs. These findings underscore urgent needs for proactive and adapted testing frameworks, more robust ethical design practices, safer testing environments, and deep cross-sector
collaborations. Our work uniquely integrates front-line law enforcement insights and a structured theoretical framework, including evaluating different human risk, providing crucial benchmarks and recommendations for future development and evaluation of safer generative models.