bebrws/security-8b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 30, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The bebrws/security-8b model is an 8 billion parameter cybersecurity agent, fine-tuned from Qwen/Qwen3-8B, specializing in both offensive and defensive security tasks. It is trained on 2197 verified agentic and Q&A traces, enabling it to plan, execute bash commands, interpret observations, and manage files within an autonomous security harness. This model excels at operational security tasks, demonstrating strong performance in synthetic attack/defend labs with a 32768-token context length and support for up to 131072 tokens via YaRN rope scaling.

Loading preview...

Overview

bebrws/security-8b is an 8 billion parameter cybersecurity agent, fine-tuned from Qwen/Qwen3-8B. It is specifically designed for autonomous security operations, capable of planning, executing shell commands, processing observations, and managing files. The model was trained on 2197 verified agentic and Q&A traces, distilled from advanced LLMs like GLM-5.2 and Kimi K3/K2.6, with a focus on iterative failure analysis and rejection sampling.

Key Capabilities

  • Autonomous Security Operations: Trained to operate an autonomous security harness for both offensive and defensive tasks, moving beyond simple Q&A.
  • Tool Use and Function Calling: Supports OpenAI-style JSON function schemas, allowing it to emit and process tool calls for shell execution and other functions.
  • Extended Context Length: Utilizes Qwen3 YaRN rope_scaling for serving up to 131072 tokens, with a default sequence length of 8192 tokens during training.
  • Strong Agentic Performance: Achieves a mean score of 8.0/10 with zero variance across three runs in a fixed 10-task synthetic attack/defend lab, outperforming its base model and other specialized cybersecurity models in agentic tasks.
  • General Capability Retention: MMLU spot checks indicate that its general language capabilities are statistically indistinguishable from the base Qwen3-8B model, showing no capability collapse despite specialized training.

Good For

  • Cybersecurity Research: Ideal as a research artifact for defensive and offensive security studies.
  • Agentic Security Tasks: Excels in scenarios requiring multi-step planning, execution, and interaction with a simulated environment.
  • Tool-Integrated Applications: Suitable for applications that leverage function calling for dynamic interaction with external systems or tools.

Limitations

  • Research Artifact Only: Not intended for real intrusion activity, exploit weaponization, unsupervised security decisions, or non-security tasks. Outputs require qualified human review.
  • Lab-Grade Competence: Attack-side competence is scoped to CTF/lab-grade tasks, not real-world, complex attack scenarios.

Finetunes
1 models