Llama-3.1-FoundationAI-SecurityLLM-8B-Reasoning (Foundation-Sec-8B-Reasoning) is an open-weight, 8-billion parameter instruction-tuned language model specialized for cybersecurity applications. It extends the Foundation-Sec-8B base model with instruction-following and reasoning capabilities. It leverages prior training to understand security concepts, terminology, and practices across multiple security domains. Further reasoning training enables the model to reason about problems before presenting a solution. Foundation-Sec-8B-Reasoning enables organizations to build AI-driven security tools that can be deployed locally, reducing dependency on cloud-based AI services while maintaining high performance on security-related tasks.

Model Name: Llama-3.1-FoundationAI-SecurityLLM-8B-Reasoning (Foundation-Sec-8B-Reasoning)
Model Developer: Foundation AI at Cisco
Model Card Contact: https://fdtn.ai/contact
Technical Report: arXiv:2601.21051
Model Release Date: January 28th, 2026
Supported Language(s): English
Model Architecture: Auto-regressive language model that uses an optimized transformer architecture (Meta Llama-3.1-8B backbone)
Training Objective: Instruction following and reasoning traces
Training Data Status: This is a static model trained on an offline dataset. Future versions of the tuned models will be released on updated data.
License: See NOTICE.md

Intended Use

Intended Use Cases

Foundation-Sec-8B-Reasoning is designed for security practitioners, researchers, and developers building AI-powered security workflows and applications. Foundation-Sec-8B-Reasoning is optimized for three core use case categories:

SOC Acceleration: Automating triage, summarization, case note generation, and evidence collection.
Proactive Threat Defense: Simulating attacks, prioritizing vulnerabilities, mapping TTPs, and modeling attacker behavior.
Engineering Enablement: Providing security assistance, validating configurations, assessing compliance evidence, and improving security posture.

The model is intended for local deployment in environments prioritizing data security, regulatory compliance, and operational control.

Downstream Use

Foundation-Sec-8B-Reasoning can be used directly for security-related chat use cases. Example downstream applications include:

Summarization
- Summarizing detection playbooks and incident reports
- Consolidating fragmented analyst notes into structured case summaries
Classification
- Mapping threats to MITRE ATT&CK techniques
- Prioritizing vulnerabilities based on contextual risk
- Classifying security-relevant emails and leaked file contents
Named Entity Recognition
- Extracting compliance evidence from documents
- Building network behavior profiles from technical manuals
Question & Answer
- Assisting SOC analysts with alert triage and investigation
- Responding to cloud security and software compliance queries
Reasoning and Text Generation
- Generating red-team attack plans and threat models
- Predicting attacker next steps in active investigations
- Enriching vulnerability scan results with contextual insights

For questions or assistance with fine-tuning Foundation-Sec-8B-Reasoning, please reach out to the team.

Out-of-Scope Use

The following uses are out-of-scope and are neither recommended nor intended use cases:

Generating harmful content - The model should not be used to:
- Generate malware or other malicious code
- Create phishing content or social engineering scripts
- Develop attack plans targeting specific organizations
- Design exploitation techniques for vulnerabilities without legitimate security research purposes
Critical security decisions without human oversight - The model should not be used for:
- Autonomous security decision-making without human review
- Critical infrastructure protection without expert supervision
- Final determination of security compliance without human verification
- Autonomous vulnerability remediation without testing
Legal or medical advice - The model is not qualified to provide:
- Legal advice regarding security regulations, compliance requirements, or intellectual property disputes
- Legal advice regarding security issues that would reference legal statutes, precedents, or case law necessary to provide legal advice
- Medical advice regarding health impacts of security incidents
Non-security use cases - The model is specifically optimized for cybersecurity and may not perform as well on general tasks as models trained for broader applications.
Violation of Laws or Regulations - Any use that violates applicable laws or regulations.

How to Get Started with the Model

Use the code below to get started with the model. The cookbook provides example use cases, code samples for adoption, and references.

python

# Import the required libraries
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("fdtn-ai/Foundation-Sec-8B-Reasoning")
model = AutoModelForCausalLM.from_pretrained("fdtn-ai/Foundation-Sec-8B-Reasoning")

prompt = "CVE-2015-10011 is a vulnerability about OpenDNS OpenResolve improper log output neutralization. What is the corresponding CWE?"

messages = [
    {"role": "user", "content": prompt}
]

model_inputs = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(model_inputs, return_tensors="pt", add_special_tokens=False)
output = model.generate(**inputs, temperature=0.1, max_new_tokens=1024)
resp = tokenizer.batch_decode(output)[0]
print(resp.replace(model_inputs, ""))

Serving with vLLM

For production deployments, you can serve Foundation-Sec-8B-Reasoning using vLLM. The model uses the minimax_m2 reasoning parser to properly handle reasoning traces.

bash

vllm serve "fdtn-ai/Foundation-Sec-8B-Reasoning" \
    --host 0.0.0.0 \
    --port ${PORT} \
    --tensor-parallel-size 1 \
    --max-model-len 32768 \
    --trust-remote-code \
    --reasoning-parser minimax_m2

Adjust --tensor-parallel-size based on your GPU configuration and --max-model-len based on your memory constraints.

Training and Evaluation

Training Data

Foundation-Sec-8B-Reasoning was trained on a wide variety of public and proprietary question answer/pairs for general and security-specific reasoning and instruction-following tasks.

Data cutoff: April 10th, 2025.

A more detailed description of the methodology is available in the technical report.

Training Setup

Foundation-Sec-8B-Reasoning is based on the Llama 3.1 8B architecture. Training was performed on Cisco Foundation AI’s internal compute cluster.

Key training details:

Instruction fine-tuning to follow human instructions
RLHF to align model answers to human preferences
32,768-token sequence length
Optimizer: AdamW

A more detailed description of the methodology is available in the technical report.

Evaluation

Foundation-Sec-8B-Reasoning was benchmarked on cybersecurity and general reasoning tasks, using a standardized 0-shot instruction prompting setup (temperature = 0.3).

Benchmark	Foundation-Sec-8B-Reasoning	Llama 3.1 8B	GPT-5-Nano
CTI-MCQA	0.691	0.607	0.688
CTI-RCM	0.753	0.531	0.672
CTI-VSP	0.856	0.811	0.822
CTI-Reasoning	0.411	0.335	0.431

Benchmark Overview:

CTI-MCQA: 2,500 multiple-choice questions testing cybersecurity knowledge across frameworks like MITRE ATT&CK, NIST, GDPR, and threat intelligence best practices.
CTI-RCM: 1,000 vulnerability root cause mapping examples linking CVEs to CWE categories, assessing deep understanding of security weaknesses.
CTI-VSP: A set of 1,000 CVE descriptions where models predict the CVSS v3 Base metrics and compute the overall score, with performance measured by the average absolute difference from the true scores.
IF-Eval: 541 instruction-following prompts designed for automated, reproducible assessment of LLM instruction-following capabilities.
Alpaca Eval 2: 805 single-turn prompts auto-scored by GPT-4 Turbo against a GPT-4 Turbo reference, validated with 20,000 human preference votes, and closely matching ChatBot Arena results.
CTI-Reasoning: An internal benchmark measuring the ability of the model to reason about second-degree connections between MITRE ATT&CK entities.

Key highlights:

Reasoning traces allow model to leverage test-time compute to answer queries
State-of-the-art non-RAG performance on CTI-RCM benchmark
Better or on-par performance on cyber threat intelligence benchmarks against GPT-5-Nano

For full benchmark details and evaluation methodology, please refer to the technical report.

Safety Alignment

Standard best practices were followed to align the model with general safety values. Despite the alignment, however, safe out-of-the-box performance cannot be guaranteed. Our evaluations show that while the model can achieve reasonable safety performance out-of-the-box, LlamaGuard provides much better protection against malicious requests. It is recommended to deploy this model with additional safeguards (such as LlamaGuard) and human oversight.

Model	HarmBench Performance
Llama-3.1-8B-Instruct	62.75%
Foundation-Sec-8B-Reasoning	93.00%
LlamaGuard + Foundation-Sec-8B-Reasoning	98.25%

Limitations

Foundation-Sec-8B-Reasoning has several limitations that users should be aware of:

Domain-specific knowledge limitations:
- Foundation-Sec-8B-Reasoning may not be familiar with recent vulnerabilities, exploits, or novel attack vectors or security technologies released after its training cutoff date
- Knowledge of specialized or proprietary security systems or tools may be limited
Potential biases:
- The model may reflect biases present in security literature and documentation
- The model may be trained on known attack patterns and have difficulty recognizing novel attack vectors
- Security practices and recommendations may be biased toward certain technological ecosystems
- Geographic and cultural biases in security approaches may be present
Security risks:
- The model cannot verify the identity or intentions of users
- Adversarial prompting techniques might potentially bypass safety mechanisms
- The model may unintentionally provide information that could be misused if proper prompting guardrails are not implemented
Contextual blindness:
- The model may struggle to understand the complex interrelationships between systems, users, and data in order to provide accurate context.
Technical limitations:
- Performance varies based on how security concepts are described in prompts
- May not fully understand complex, multi-step security scenarios without clear explanation
- Cannot access external systems or actively scan environments
- Cannot independently verify factual accuracy of its outputs
Ethical considerations:
- Dual-use nature of security knowledge requires careful consideration of appropriate use cases

Recommendations

To address the limitations of Foundation-Sec-8B-Reasoning, we recommend:

Human oversight:
- Always have qualified security professionals review model outputs before implementation
- Use the model as an assistive tool rather than a replacement for expert human judgment
- Implement a human-in-the-loop approach for security-critical applications
System design safeguards:
- Implement additional validation layers for applications built with this model
- Consider architectural constraints that limit the model’s ability to perform potentially harmful actions (excessive agency)
- Deploy the model in environments with appropriate access controls
Prompt engineering:
- Use carefully designed prompts that encourage ethical security practices
- Include explicit instructions regarding responsible disclosure and ethical hacking principles
- Structure interactions to minimize the risk of inadvertently harmful outputs
Knowledge supplementation:
- Supplement the model with up-to-date security feeds and databases
- Implement retrieval-augmented generation for current threat intelligence sources
Usage policies:
- Develop and enforce clear acceptable use policies for applications using this model
- Implement monitoring and auditing for high-risk applications
- Create documentation for end users about the model’s limitations

Citation

bibtex

@misc{yang2026foundation-sec-8b-reasoning,
      title={Llama-3.1-FoundationAI-SecurityLLM-Reasoning-8B Technical Report}, 
      author={Zhuoran Yang and Ed Li and Jianliang He and Aman Priyanshu and Baturay Saglam and Paul Kassianik and Sajana Weerawardhena and Anu Vellore and Blaine Nelson and Neusha Javidnia and Arthur Goldblatt and Fraser Burch and Avi Zohary and Assaf Eisenman and Mahdi Sabbaghi and Supriti Vijay and Rahim Dharssi and Dhruv Kedia and Kojin Oshiba and Yaron Singer and Amin Karbasi},
      year={2026},
      eprint={2601.21051},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://huggingface.co/papers/2601.21051}
}