UC Berkeley study finds fine-tuning can quietly break AI privacy safeguards

A UC Berkeley study found that fine-tuning AI models can quietly erode their privacy safeguards, undermining the standard safety evaluations that businesses rely on. The research shows that models which pass privacy checks in their base form may leak sensitive training data after customization, creating hidden compliance and data-protection risks for companies deploying tailored AI. The findings matter because fine-tuning is a routine step for enterprises adapting models to proprietary data, yet privacy testing typically happens before, not after, that process. What to watch next is whether AI developers and regulators extend privacy audits to fine-tuned variants and set post-tuning evaluation standards.