A new Python package called reskSecure has been developed to act as a firewall for large language models, aiming to prevent jailbreaks and prompt injections. Unlike traditional methods that rely on prompt engineering or post-generation filtering, reskSecure intervenes directly within the model's token generation loop. It utilizes a 64-bit permission bitmask to control what the LLM is allowed to output, setting logits to negative infinity for hard blocks or applying penalties for bias, thereby enhancing security. AI
IMPACT This tool could significantly improve the security of LLM applications by preventing unauthorized actions and data leaks.
RANK_REASON The item describes a new software package that enhances LLM security, fitting the definition of a tool.
- Aho–Corasick algorithm
- AutoModelForCausalLM
- AutoTokenizer
- Bitmask
- mistralai/Mistral-7B-v0.1
- Python
- resklogits
- reskSecure
- transformers
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →