pybbb/Qwen3.6-27B-SecOPD
pybbb/Qwen3.6-27B-SecOPD is a 27 billion parameter causal language model, based on the Qwen3.6 architecture, developed by pybbb. This model is specifically fine-tuned using the SecOPD (Mitigating Adaptive Prompt Injections by On-Policy Distillation) method to significantly enhance robustness against adaptive indirect prompt injections. It achieves a 9.0% PISmith ASR, substantially lower than undefended models, while preserving utility. The model is designed for secure inference in applications requiring explicit trusted/untrusted input boundaries, such as prompt injection research and red-team evaluations.
Loading preview...
Overview
pybbb/Qwen3.6-27B-SecOPD is a 27 billion parameter causal language model built upon the Qwen3.6 architecture, developed by pybbb. Its core innovation lies in the application of SecOPD (Mitigating Adaptive Prompt Injections by On-Policy Distillation), a training method that uses clean-context token-level supervision to improve robustness against adaptive indirect prompt injections. This model is designed to maintain high utility while significantly reducing vulnerability to prompt injection attacks.
Key Capabilities & Performance
- Enhanced Prompt Injection Defense: Achieves a remarkably low 9.0% PISmith ASR (Attack Success Rate) compared to 97.9% for undefended models, and 0.2% SEP Basic Adaptive ASR. This makes it highly resistant to various forms of indirect prompt injection.
- Utility Preservation: Demonstrates strong performance across utility benchmarks, including 80.1% on AlpacaEval2, 88.6% on SEP Utility, and 84.1% on MMLU-Pro, indicating that its security enhancements do not significantly degrade general model capabilities.
- Explicit Trust Boundary: Designed to process trusted instructions in the
userrole and untrusted content (e.g., documents, tool returns, images) in theinputrole, enforcing a clear security boundary. - Vision Encoder: Inherits a vision encoder from its base Qwen3.6 architecture, allowing for multimodal input processing.
Intended Use Cases
- Prompt Injection Research: Ideal for studying and evaluating prompt injection vulnerabilities and defenses.
- Red-Team Evaluation: Suitable for red-teaming LLM applications to identify security weaknesses.
- Secure Application Prototyping: Useful for developing applications where maintaining an explicit trusted/untrusted input boundary is critical for security, such as those handling user-generated content or external data sources.
Limitations
While significantly more robust, prompt injection is not entirely solved. The model relies on the host application to correctly identify and separate trusted and untrusted inputs. It is a model-level defense against indirect prompt injection and does not replace jailbreak defenses or system-level authorization.