Cisco and former Google and Meta engineers develop cybersecurity-specific large language models

Cisco’s Foundation Gen AI group released an open-source, cybersecurity-trained Llama 3 model, optimized to run efficiently on a single GPU, aiming to improve threat detection and response in enterprise security operations.

Cisco’s new Foundation Gen AI group, led by former Harvard professor Yaron Singer, has trained a cybersecurity-focused version of Meta’s Llama 3 large language model.

The model, based on 5 billion tokens distilled from 200 billion cybersecurity-relevant tokens, is open-sourced with open weights for public use. Designed for speed and efficiency, it can run on a single Nvidia A100 GPU, making it cost-effective for enterprises. Cisco plans to integrate the model into its extended detection and response (XDR) products.

This specialized LLM addresses cybersecurity’s dynamic, non-standard data challenges, offering organizations the flexibility to fine-tune it for their unique environments.

OpenAI