Cisco creating new security model using 30 years of data describing cyber-dramas and saves
- Reference: 1762757812
- News link: https://www.theregister.co.uk/2025/11/10/cisco_new_17bn_parameter_model/
- Source link:
The networking giant’s current hero model is [1]Foundation-Sec-8b , an eight-billion parameter effort that it uses in several products, and suggests for use on tasks such as automated processing of security alerts, reviewing code to find potential security issues, and suggesting workflows that prioritize security.
At the Asia-Pacific edition of its Cisco Live event in Melbourne today, Cisco’s senior vice president and chief product officer for security, Raj Chopra, said Cisco is working on a new model that will use 17 billion parameters.
[2]
As was the case with Foundation-Sec-8B, Cisco will train the new model on threat information, incident summaries, red team playbooks, and other info it’s collected. Chopra said this new model will include data from Cisco’s Talos threat intelligence team, which has 30 years’ worth of data to contribute.
[3]Cisco Borgs all its management tools into a single Cloud Control console
[4]Cisco: Most companies don't know what they're doing with AI
[5]Hungry hyperscalers boosted Cisco's AI sales by a cool billion bucks
[6]Senator presses Cisco over firewall flaws that burned US agency
Chopra told The Register this new model is not a successor to Foundation-Sec-8B, but instead a fresh effort because Cisco’s goal is to build a model that can both detect threats and advise on steps to address them. “And for that we need an expanded model,” he said, adding that it will appear “after Christmas, but not far after that.”
The senior veep said Cisco is working on “a whole phalanx” of new models and AI initiatives but didn’t offer any details. He did, however, point to the late October [7]release of an update to the SecureBERT model, a tool designed as an aide to security professionals.
[8]
Chopra said Cisco helped the model’s developers to “enhance performance manyfold” in the new release.
Cisco develops LLMs because it believes organizations need a mix of generic security data and info describing their own affairs to effectively use AI in their defenses. The company's models are open source, but its AI efforts are not entirely altruistic as it uses them to power some of its products. Cisco also suggests its Splunk tools as the best way for customers to analyze their own data. ®
Get our [9]Tech Resources
[1] https://huggingface.co/collections/fdtn-ai/foundation-sec-8b
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_security/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aRHFySQViTQoRAj5W4VwuwAAAEM&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://www.theregister.com/2025/06/10/cisco_live_cloud_control_news/
[4] https://www.theregister.com/2025/10/15/cisco_ai_readiness_index/
[5] https://www.theregister.com/2025/08/14/cisco_q4_2025/
[6] https://www.theregister.com/2025/10/16/cisco_senate_scrutiny/
[7] https://blogs.cisco.com/ai/securebert-2-0-ciscos-next-gen-ai-model-powering-cybersecurity-applications
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_security/front&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aRHFySQViTQoRAj5W4VwuwAAAEM&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[9] https://whitepapers.theregister.com/
Re: From 8 to 17 billion
https://huggingface.co/fdtn-ai says the 8b model and it's underlying Llama 3.1 model were already trained at BF16. I doubt they've gone lower for the new model.
I'd guess the new one is a Llama 4 variant which is also BF16 but has a mixture-of-experts architecture. If Cisco only used one expert, which is perhaps/probably sensible for a specialised task(?), it'd be 17b paramaters. Or just possibly it has multiple experts and they're underselling it a bit.
Open Source?
Wow, kind of impressed with that and may actually get me to have a deeper look at it.
Again, I have to state that I do not believe that LLMs are ready for the big time wrt, well, doing almost anything, but a specifically trained model such as this that can give advice on potential solutions is a better assistant than something like ChatGPT (IMHO).
new "security model" ?
I was thinking from the headline just implementing an existing security model correctly would a giant leap forward.... but then reading the subhead I realized that this "model" actually referred to more raw AI sewerage† to further contaminate the pool.
† alias liquid shit.
From 8 to 17 billion
I wonder if they switched from 32-bit precision to 16-bit precision ?
Because I doubt that they doubled the amount of RAM.