A new auditing technique identifies dangerous AI model adaptations without generating illegal content, offering platforms a scalable tool to improve online child safety efforts.
Massachusetts Institute of Technology (MIT) researchers, working with Thorn, have developed a new auditing technique that identifies generative AI models adapted to produce illegal child sexual abuse material (CSAM) without requiring the models to generate any harmful content. The method, presented at the Trustworthy AI for Good workshop at the International Conference on Machine Learning, achieved 100% accuracy during testing on specialised model variants.
The approach addresses a growing challenge as open-source AI models become increasingly available and can be fine-tuned for legitimate tasks or misused to create harmful content. Traditional auditing methods require prompting models and examining their outputs, but this is unsuitable for CSAM because generating such material is illegal in many jurisdictions and can expose human reviewers to distressing content.
Instead of analysing outputs, the researchers inspect modifications introduced during fine-tuning, known as low-rank adaptation (LoRA) adapters. Using a process called Gaussian probing, the system feeds models random data and examines how the information is processed internally. This allows auditors to determine whether a model has been adapted for harmful capabilities without producing any images or illegal outputs.
During evaluation, the technique successfully distinguished models modified to generate CSAM from safe and other specialised models. The researchers say the method is scalable, relatively inexpensive to implement and well suited for hosting platforms that review thousands of uploaded AI model variants each month. It could help detect unsafe models before they are widely distributed or shared online.
The team also believes the technique is more robust than many existing auditing approaches because malicious actors would need to substantially alter a model’s internal structure to evade detection. Future work will expand testing across a broader range of AI models and investigate whether similar methods can identify harmful capabilities already embedded within base models before they are fine-tuned for misuse.
