ישראל מירסקי

Large Language Models (LLMs) are increasingly being used to automate deepfakes calls and perform social engineering attacks. While current research focuses on designing robust safeguards, there is a critical gap in developing defenses capable of generalizing to a variety of models -and to models that have no safeguards from the start (e.g., finetuned source models). In this work, we explore "token mines"- a sequence of tokens that, if consumed by an LLM, corrupt its hidden state and cause the model to collapse. These token mines can be deployed in a conversation to corrupt a model and thus prove to the victim that he or she is talking with an AI. To create these mines, we optimize a model's output to maximize entropy, resulting in a sequence that not only collapses the base model but also other models as well (i.e., transfers). We evaluate these mines across seven open models spanning GPT-2, LLaMA, BLOOM, Qwen, and Phi architectures with vocabularies ranging from 32K to 250K tokens. Our optimized token sequences achieve 89-96% normalized self-optimization entropy and induce degenerate outputs - including repetition loops, garbage token floods, and cross-lingual artifacts - with over 95% corruption rate in cross-model transfer tests. These findings demonstrate that long-tail vulnerabilities are broadly shared across architectures, suggesting that Token Mines can serve as a viable "universal barrier"for proactively neutralizing unauthorized agents.