Cisco is launching a new initiative today at the RSAC 2025 Conference: Foundation AI. This new team within Cisco says it will focus on developing cutting-edge AI technology. One of its first achievements is a new AI language model that focuses specifically on cybersecurity.
Foundation AI can be seen as a continuation of the ambitions of Robust Intelligence, a company Cisco acquired last year. It is therefore no coincidence that the new team is led by Yaron Singer, university lecturer and co-founder of Robust Intelligence. He is now VP AI & Security at Cisco and in that capacity also responsible for Foundation AI. As the name suggests, Foundation AI is all about the basics of AI and about securing AI, which requires a lot of scientific research. That is why the team consists mainly of scientists, but also developers and other profiles.
An important part of Foundation AI is that it is an open-source initiative. According to Singer, everything they do and build will eventually be made available to the community and, of course, to customers.
Standard LLMs are not sufficient for security purposes
LLMs have been used for cybersecurity purposes for some time. However, Singer sees that the impact of LLMs on cybersecurity is much smaller than on other segments, such as programming. This is mainly because standard LLMs are not built for security and are therefore not really suitable for it, partly due to the very specific language used in cybersecurity. Existing closed-source models are also difficult or impossible to adapt.
Foundation AI introduces open-source security model
Foundation AI has set itself the goal of solving this problem, or at least contributing to its solution. Singer claims that Foundation AI has developed the first “Security AI Reasoning Model.” This model is, of course, open source.
In a session we attended, Singer explained how the new model came about. It is a model with 8 billion parameters. This model has been trained exclusively on cybersecurity data that also comes from open-source sources. It involves 5 billion tokens, which in turn have been distilled from 200 billion tokens. The idea is that this makes it a highly concentrated model containing only relevant data. This yields much better results than a much larger but less specialized dataset can achieve. It can run on a single Nvidia A100 GPU, making it accessible to many more customers.
In addition to the specific characteristics of the new AI model, Singer believes it is also important that the model is easy to fine-tune. Even a highly specialized model will not provide 100% coverage for every organization. That is why Foundation AI has paid particular attention to this aspect.
How is this different from Trend Cybertron?
The establishment of Foundation AI by Cisco and the introduction of the new model is undoubtedly a good move. However, it als have us a sense of déjà vu. That is, it reminded us of an announcement by Trend Micro earlier this year. At first glance, Trend Cybertron seems to offer virtually the same thing. It is also based on Llama 3 (3.1, to be precise), uses 8 billion parameters, and Trend Micro has now also made the model open source.
The basis for Cybertron and the new model developed by Foundation AI is therefore identical. In itself, this is not surprising. That’s quite common when something is built on open-source technology. Based on the information we can find online, however, Cisco seems to have taken it a few steps further than Trend Micro.
Trend has made a “select number of models” available to the open source community and seems to be focusing primarily on the bigger Trend Cybertron, which is part of its own Vision One platform. The models that Trend makes available can be found via the Nvidia NIM catalog, Huggingface, and Github. The reasoning component that Cisco has invested time in seems to have been left to other parties by Trend. We also cannot find any information about the distillation of the token set for Trend Cybertron, which is a crucial component in terms of accuracy.
The beginning of a new trend in AI models?
At this point, based on the information currently available, we find Cisco’s claim that it is the first with an open-source “Security AI Reasoning Model” just about tenable. In itself, who is first and who is second is not very interesting anyway. We expect more parties to follow suit anyway. Then it will no longer matter who exactly was first. If anything, it is a good sign that we already have two of these models on the market, and that they have been open-sourced by companies who haven’t been known historically as being very open-source minded. That also tells us that there’s real potential in this market, but that’s something for another time.
Read also: Cisco AI Defense enables the secure deployment of AI