Logit-Level Filtering Proposed as Stronger Defense Against LLM Jailbreaks

A new open-source tool called resk-logits aims to address security gaps in large language models by intercepting token probability distributions before text is generated, rather than scanning outputs after the fact. Traditional guardrails, regex filters, and audits operate post-sampling, meaning a jailbreak has already occurred at the logit level by the time they detect it. The tool uses Aho-Corasick pattern matching on the GPU to suppress harmful token sequences proactively, with claimed processing speeds under one millisecond for over 10,000 patterns. Developed by Resk Security, the library is available on GitHub and PyPI. The developers argue that while audits and output filters remain useful, true LLM security requires intervening at the point where token decisions are actually made.