pybbb/Qwen3.6-27B-SecOPD

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

pybbb/Qwen3.6-27B-SecOPD is a 27 billion parameter prompt-injection defense model built on Qwen3.6-27B, developed by pybbb. It is specifically trained using Secure On-Policy Distillation (SecOPD) to mitigate adaptive prompt injections by separating external data from trusted user instructions. This model excels at enhancing the security of LLM applications against malicious inputs, demonstrating significantly reduced attack success rates while maintaining utility.

Loading preview...

Qwen3.6-27B-SecOPD: Prompt Injection Defense

Qwen3.6-27B-SecOPD is a specialized 27 billion parameter language model, fine-tuned from Qwen3.6-27B, designed to robustly defend against prompt injection attacks. Developed by pybbb, this model utilizes Secure On-Policy Distillation (SecOPD), a novel training method that provides token-level feedback from a frozen teacher model to the student, conditioned on clean inputs. A key feature is its use of a dedicated input role to clearly distinguish external, untrusted data (like documents or tool results) from trusted user instructions.

Key Capabilities

  • Robust Prompt Injection Defense: Achieves significantly lower Attack Success Rates (ASR) compared to undefended models and other defense mechanisms like Meta-SecAlign and GRPO, particularly on benchmarks like SEP Static (1.3% ASR) and SEP Basic Adaptive (0.2% ASR).
  • Maintained Utility: Despite its security focus, the model largely preserves utility across various benchmarks, including AlpacaEval2 (80.1%), SEP Utility (88.6%), and MMLU-Pro (84.1%), ensuring it remains effective for general tasks.
  • Dedicated Input Role: Employs a unique input role in its chat template, allowing developers to explicitly separate potentially malicious external content from safe user prompts, enhancing control and security.

Good For

  • Applications requiring high security against prompt injections: Ideal for systems where external data or user-provided content could be manipulated to bypass safety mechanisms or extract sensitive information.
  • Integrating untrusted data sources: Safely process retrieved passages, tool outputs, or user-generated content without compromising the model's intended behavior.
  • Developers building secure LLM agents: Provides a strong foundation for building agents that can resist adaptive and straightforward injection attempts, as detailed in the SecOPD paper.

Finetunes
1 models