Researchers have developed a method to prevent large language models from recovering forgotten information through external tools. The method, called Agentic Tool Unlearning (ATU), consists of two stages: first, it suppresses direct recall by unlearning model parameters, and second, it penalizes target-seeking behavior in simulated environments. Experiments show that ATU achieves a better balance between forgetting and retaining useful knowledge than previous methods. --- Why it matters: This matters to engineers working on large language models because current unlearning methods can be bypassed by external tools, undermining their effectiveness. ATU's ability to prevent tool-mediated recovery is crucial for deploying LLMs as reliable agents in real-world applications. Source: https://arxiv.org/abs/2608.21544