Cisco is proud to mark a serious milestone in its AI journey. Our newly launched Basis AI group—shaped to convey world-class, domain-specific AI infrastructure to the cybersecurity house—is proud to introduce its first launch: Llama-3.1-FoundationAI-SecurityLLM-base-8B (Basis-sec-8b), an 8-billion parameter, open-weight Massive Language Mannequin (LLM) purpose-built for safety.
This mannequin combines deep area experience with the flexibleness and management wanted for safe, real-world deployment. With Basis-sec-8B, groups can construct, fine-tune, and deploy AI-native workflows throughout the safety lifecycle. This mannequin is designed to assist safety groups assume sooner, act with precision, and scale operations with out compromise.
Cybersecurity Purposes Require Goal-Constructed, open, Small Fashions
Safety groups at the moment face unprecedented stress. Threats are rising sooner, assaults are extra complicated, and conventional instruments aren’t sufficient. AI is rising as a crucial power multiplier—however general-purpose fashions typically fall quick for real-world safety workflows.
Prompting and retrieval pipelines can’t meet the precision, consistency, and high quality calls for of safety operations. Nice-tuning closed fashions brings its personal challenges: excessive prices, restricted management, and deployment limitations—particularly the place knowledge privateness, regulatory compliance, and in-environment internet hosting are non-negotiable. In the meantime, most open fashions lack the domain-specific information safety groups want.
“Basis-sec-8b is a foundational step towards constructing AI-native safety techniques—instruments that don’t simply course of knowledge however actually perceive the safety area. This mannequin provides safety groups in every single place a robust constructing block to speed up protection, cut back fatigue, and achieve readability in complicated risk environments.”
Jeetu Patel, Chief Product Officer, Cisco
A Goal-Constructed Mannequin for Safety
Basis-sec-8b isn’t a general-purpose mannequin retrofitted for safety. It was designed from the bottom as much as perceive the language, logic, and workflows of cybersecurity.
Constructed with Llama, utilizing the Llama 3.1 8B framework, foundation-sec-8b has been pre-trained on a dataset meticulously constructed in-house by Cisco Basis AI. This dataset was curated particularly to seize the real-world information and context that safety professionals work with day-after-day. It contains, however shouldn’t be restricted to:
- Vulnerability databases and risk conduct mappings (e.g., CVEs, CWEs, MITRE ATT&CK)
- Menace intelligence reviews, pink group playbooks, and real-world incident summaries
- Safety tooling documentation throughout cloud, id, and infrastructure domains
- Compliance references and safe growth practices (e.g., NIST, OWASP, safe coding guides)
This area alignment makes foundation-sec-8b instantly efficient throughout analyst workflows—delivering deeper insights, fewer hallucinations, and dramatically decrease response instances in comparison with generalist fashions.
Efficiency That Punches Above Its Weight
Regardless of its 8B parameter measurement, foundation-sec-8b outcomes are comparable with fashions almost 10x bigger. On core cybersecurity benchmarks, it outperforms Llama 3.1 8B and even matches or exceeds Llama 3.1 70B, providing highly effective inference at a fraction of the associated fee and infrastructure burden.
| Benchmark | Mannequin Basis-sec-8b |
Mannequin Llama-3.1-8b |
Mannequin Llama-3.1 70B |
|---|---|---|---|
| CTI-MCQA | 67.39 | 64.14 | 68.23 |
| CTI-RCM | 75.26 | 66.43 | 72.66 |
Importantly, foundation-sec-8b maintains sturdy basic language efficiency, staying on par with Llama fashions on MMLU benchmarks. This implies it might fluidly clarify risk conduct, help long-form investigation narratives, and reply naturally in chat-based workflows—with out compromising its safety focus.
Its benefits develop much more obvious with fine-tuning. In downstream safety duties—equivalent to extracting MITRE ATT&CK strategies from unstructured risk reviews—a fine-tuned foundation-sec-8b mannequin considerably outperformed fine-tuning the same-sized Llama mannequin. This reinforces the worth of foundation-sec-8b as each a foundational mannequin and a high-impact element for embedded safety workflows.
Sensible Purposes—and Constructed to Match Your Atmosphere
Basis-sec-8b isn’t simply performant—it’s versatile. Being open-weight and constructed on the Llama framework makes it simple to fine-tune or prolong with your individual telemetry, detection guidelines, or risk intel. Whether or not you’re embedding it into present instruments or constructing totally new workflows, foundation-sec-8b adapts to your group’s distinctive wants.
Safety groups can apply foundation-sec-8b throughout the safety lifecycle:
- SOC acceleration – Automate alert triage, summarize incidents, and help with investigations
- Proactive risk protection – Simulate assaults, prioritize vulnerabilities, and mannequin threats tailor-made to your infrastructure
- Engineering enablement – Ship AI-assisted code opinions, validate configurations, and assess compliance proof with context out of your surroundings
- Customized integration – Nice-tune the mannequin to replicate your safety posture, language, and detection methods—making certain precision in real-world deployments
“We have now been working with the Cisco Basis AI group exploring the best way to embed foundation-sec-8b throughout Cisco’s inner safety stack. We’re very excited in regards to the capabilities of an AI mannequin that actually understands our area and may scale throughout use circumstances.”
Omar Santos, Distinguished Engineer, Cisco Safety & Belief Group (S&TO)
From powering agentic assistant workflows to helping builders in writing secure-by-default code, foundation-sec-8b is a basis you may construct on.
Publicly Obtainable, With Privateness and Management in Thoughts
In cybersecurity, belief is non-negotiable. That’s why foundation-sec-8b is open-weight, with weights and tokenizer launched below a permissive license.
This offers organizations:
- Full management over deployment – Run the mannequin on-prem, in air-gapped environments, or inside safe cloud enclaves
- Compliance confidence – Preserve delicate knowledge native; no pressured inference APIs or third-party sharing
- Customization freedom – Adapt and prolong the mannequin structure or coaching pipeline to satisfy distinctive safety and privateness wants
Basis-sec-8b is now accessible for obtain on Hugging Face.
You may also learn the total technical particulars in our analysis paper.
Within the close to future, we additionally plan to open supply the coaching pipeline used to construct the mannequin. For extra on the information building and coaching course of, discuss with the analysis paper.
Simply the Starting
Basis-sec-8b is the first launch from Cisco Basis AI, a devoted group constructing the infrastructure that may drive the subsequent decade of cybersecurity innovation. This mannequin lays the groundwork for a future the place AI isn’t only a bolt-on characteristic—however a foundational functionality embedded throughout all the safety lifecycle.
And we’re simply getting began.
Over the approaching months, Cisco Basis AI will probably be releasing:
- A cybersecurity reasoning mannequin that brings explainability and deeper evaluation to complicated safety workflows
- A brand new benchmark suite designed to guage AI fashions on real-world, practitioner-defined safety duties
- Further instruments and parts that assist groups fine-tune, operationalize, and embed AI safely and successfully into their safety stacks
Basis-sec-8b opens the door. What comes subsequent will reshape how defenders construct, reply, and lead with AI.
Basis-sec-8b is stay. It’s open. And it’s able to defend.
Discover the mannequin, be a part of the group, and assist form the subsequent technology of AI-powered cybersecurity.
We’d love to listen to what you assume. Ask a Query, Remark Under, and Keep Related with Cisco Safety on social!
Cisco Safety Social Handles
Share:
