New open-source tool blocks LLM jailbreaks at logit level

A new open-source tool called resk-logits has been released to enhance LLM safety by blocking harmful content at the logit level, before a token is sampled. This GPU-accelerated processor uses an Aho-Corasick algorithm to scan the entire vocabulary for malicious patterns, setting the logits of matching tokens to negative infinity for a hard block or applying a penalty for a "shadow ban" mode. Developed with both CUDA and Metal backends, it offers sub-millisecond overhead and is compatible with any Hugging Face AutoModelForCausalLM. AI

IMPACT Enhances LLM safety by providing a low-latency, architecture-level blocking mechanism for harmful content.

RANK_REASON New open-source tool release for LLM safety.

Read on dev.to — LLM tag →

• Aho–Corasick algorithm
• AutoModelForCausalLM
• CUDA
• GitHub
• Hugging Face
• LLM
• Metal
• Python Package Index
• resk-logits
• RTX 4090

• product
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →