Large Language Models (LLMs) that provide access to output logits are susceptible to a sophisticated attack known as "coercive interrogation." This method enables the extraction of concealed and potentially harmful knowledge embedded within low-ranked tokens. Unlike traditional prompt-based attacks, this technique does not necessitate specially crafted prompts. Instead, it systematically compels the LLM to select and produce low-probability tokens at strategic points within the response sequence. This approach unveils toxic content that the model typically suppresses. Mitigation steps: **For AI Developers:** * Restrict API access to output logits, ensuring users receive only the final, filtered output. * Implement enhanced filtering mechanisms to identify and remove toxic content from responses, utilizing advanced techniques beyond simple keyword blocking. **For Model Trainers/Fine-tuners:** * Develop improved model alignment techniques that are resistant to coercion and effectively suppress harmful knowledge. * Investigate methods for model unlearning to remove harmful knowledge from the training data or model weights.