A collaborative research team from MIT and the child safety nonprofit Thorn has developed a novel AI auditing method designed to detect models fine-tuned to generate illegal child sexual abuse material without producing harmful outputs. The initiative addresses a critical gap in AI safety as generative models become increasingly accessible and easily modified through low-rank adaptation algorithms. According to the National Center for Missing and Exploited Children, reports of AI-generated CSAM surged to over 1.5 million in 2025, a sharp increase from 67,000 the previous year. Traditional safety evaluations rely on prompting models to generate prohibited content and manually inspecting the results, a practice that is both illegal under U.S. law and psychologically taxing for human evaluators. To bypass these constraints, the researchers introduced a non-generative auditing framework that examines the internal modifications of fine-tuned models rather than their final outputs. Led by MIT graduate student Vinith Suriyakumar alongside associate professors Ashia Wilson and Marzyeh Ghassemi, the team analyzed low-rank adaptation layers using a technique called Gaussian probing. By feeding the model random data points and measuring how its internal neural layers manipulate those inputs, the method captures the computational fingerprints left by harmful fine-tuning processes. The approach never prompts the model or generates images, effectively eliminating legal and ethical barriers to testing. In controlled trials across three categories of generative models, the Gaussian probing technique achieved 100 percent accuracy in identifying variants specialized for producing CSAM. The method also successfully distinguished between adaptations designed for other harmful imagery and those trained on benign datasets. Because the evaluation targets structural modifications rather than surface-level outputs, it remains robust against attempts to obfuscate malicious capabilities. The researchers emphasized that the technique is computationally efficient and scalable, making it suitable for deployment by model hosting platforms and law enforcement agencies tasked with filtering thousands of newly published AI variants monthly. The findings were published as a spotlight presentation at the Trustworthy AI for Good workshop held alongside the International Conference on Machine Learning. The project involved researchers from MIT, Boston University, and Thorn, with institutional support from the Bridgewater AIA Labs Research Fellowship. By transforming an area previously considered an unmeasurable blind spot into a quantifiable safety metric, the work establishes a new standard for pre-deployment AI audits. The team plans to expand testing to a broader range of model architectures and investigate whether the probing methodology can identify harmful precursors within base models before fine-tuning occurs. This development marks a significant step toward mitigating the proliferation of illegal AI-generated content while preserving the constructive applications of open-source generative systems.