A new initiative has emerged within Cisco, spearheaded by their recently established Foundation AI research group, which has undertaken the task of training Meta’s Llama 3 large language model (LLM) specifically on cybersecurity data. This important model is set to be released as open source, providing publicly accessible open weights.
The Foundation AI group was introduced to the public on Monday and is headed by Yaron Singer, who previously served as a professor of computer science and advanced mathematics at Harvard University. Additionally, he holds the position of CEO and co-founder of Robust Intelligence, a company that Cisco acquired in 2024. Currently serving as the vice president of AI and security at Cisco, Singer has strategically gathered a team of engineers from both Meta and Google to collaborate on training this specialized cybersecurity LLM.
This release marks a significant milestone in the open-source landscape, as the model will be available for anyone to download, inspect, and fine-tune. While its parameters will be accessible, the source code and underlying data will remain proprietary. Cisco is planning to integrate this advanced model with AI agents in its extended detection and response product line. Furthermore, they recently introduced AI agents designed for attack verification, automated forensics, and a visualization tool known as Attack Storyboard, utilizing various LLMs.
Singer noted the unique nature of cybersecurity data, emphasizing that it often consists of specialized languages rather than conventional natural language. “Cybersecurity data, by its nature, is not necessarily natural language — it’s often bespoke languages,” he explained. “It’s dynamic, so threats and vulnerabilities get updated frequently, and all that makes existing AI tools that we have right now for cybersecurity not sufficient for the SOC [Security Operations Center] to adopt them.”
The Foundation AI project has meticulously distilled open-source data from an extensive collection of 200 billion tokens down to 5 billion tokens that are most pertinent to cybersecurity. This careful curation aims to enhance the model’s efficiency and performance. Although Cisco has not disclosed precise benchmark numbers, Singer asserted that the LLM is smaller than most foundational models and is capable of operating on a single Nvidia A100 GPU on-premises.
Industry analysts, like Andy Thurai from The Field CTO, highlighted the potential for IT organizations to personalize the model further by integrating their own retrieval-augmented generation data, thus enhancing its applicability within specific environments. Thurai remarked on the limitations of current general-purpose LLMs, noting, “Current general-purpose LLMs are mostly used for security-to-human-understanding translation with varying success, unlike this.” He emphasized that the capability to run on a single A100 GPU is particularly impressive, allowing even budget-conscious customers to utilize this resource without facing exorbitant costs associated with larger LLMs.
In a broader context, the dissemination of specialized LLMs is seen as a natural evolution in the field of artificial intelligence. Adrian Sanabria, an independent security consultant, pointed out that the growing demand for specialized models stems from an expectation of limited returns on more generalized foundational models. “Increasing specialization of LLMs is an expected evolution now that initial foundation models have been established,” he said, suggesting that reasoning models in agentic architectures can allocate tasks efficiently to the most suitable model, API, or service.
However, Sanabria also warned that the increasing reliance on specialized AI agents, including those utilizing LLMs trained for cybersecurity, may confront scalability challenges as their deployment becomes more widespread. Cost remains a significant factor, as an individual A100 GPU can cost around $8,000, and agentic AI systems tend to consume more energy than traditional IT workloads. Moreover, cybersecurity startups like Panther have indicated that security operations centers (SOCs) typically deal with over 4,000 alerts daily, translating to an overwhelming number of alerts to process.
Sanabria highlighted the urgency of effective alert management amid the demands placed on SOCs: “If each alert is taking AI SOC agents three minutes in the best-case scenario, that’s a limit of 480 alerts per day, assuming their estimate is for a single GPU.” Although this performance is notably swifter than the average human analyst, who may take 20 to 40 minutes per alert, the challenge of using AI for rapid alert processing remains evident.
Ultimately, while agentic AI offers promising advantages, it alone will not eliminate SOC alert fatigue. Sanabria emphasized the importance of selectivity concerning which alerts should be addressed by AI due to its inherent costs and limitations. He pointed out that existing AI-based tools can still perform effective event correlation and alert reduction without relying on LLMs, with some managed security service providers automating tasks at scale without utilizing LLM technology.
The potential risks of releasing a cybersecurity LLM as open source have also been raised by analysts. Thurai expressed concern that malicious actors could exploit such models for scanning vulnerabilities and launching attacks. However, he did note that if the Cisco Foundation AI model performs as intended, it could significantly aid enterprises in identifying vulnerabilities, particularly during red-teaming exercises.
In summary, the introduction of a cybersecurity-focused LLM by Cisco represents a noteworthy advancement in the field of AI-driven cybersecurity tools. With the ability to enhance operational efficiency for security operations centers, this initiative is poised to be a game-changer in the ongoing battle against cyber threats. As organizations continue to grapple with a challenging threat landscape, innovations like these may prove crucial in fortifying defenses and navigating the complexities of modern cybersecurity.