Imagine training a powerful AI on your company's data to create an email assistant, but that data also has restricted information. How do you prevent the AI from accidentally revealing financial reports to the marketing team, while still letting it craft marketing emails that draw upon company-wide data for peak effectiveness? This is the core challenge of access control for Large Language Models (LLMs). Researchers at Ben Gurion University have developed a clever solution called DOMBA – Double Model Balancing – that elegantly sidesteps this problem. DOMBA’s secret lies in training two separate AI models. Each model learns from different parts of your company’s data, according to the access restrictions in place. One model might learn from sales and finance, while the other focuses on marketing, HR, and product. The magic happens during text generation: DOMBA doesn’t just rely on one model. Instead, it uses a special averaging method that blends the outputs of *both* models to determine the probability of each word in the generated text. This "min-bounded" averaging prioritizes agreement between the models. If one model is highly confident about a word or phrase, while the other is uncertain, it indicates specialized information. This word is much less likely to be included in the final generated output. This ingenious approach keeps the LLM from revealing information only one model "knows," thus safeguarding restricted data. The team behind DOMBA introduced some clever evaluation metrics: 'exposure,' to measure how much sensitive information leaks out; ‘secret perplexity,’ assessing how well the AI handles sensitive phrases; and a ‘secret inference attack’ that mimics how a hacker might try to extract information. Results across movie reviews and recipes demonstrate that DOMBA provides a solid balance between utility and security. It doesn't perfectly shield *all* restricted information, but achieves a robust, practical level of protection while keeping the LLM useful. Future development of DOMBA holds exciting possibilities. The min-bounded averaging core is an elegant solution for managing sensitive information in LLMs, whether for access control or general data privacy. It also allows for more training epochs and a more extensive dataset without the same risk of overfitting as other solutions, while still producing an effective balance between utility and privacy.