opena2a/nanomind-security-analyst
The opena2a/nanomind-security-analyst is a 2 billion parameter generative threat analysis model, fine-tuned from Qwen3-1.7B by opena2a, designed for structured security analysis of AI agent artifacts. It excels at identifying and classifying security threats within npm packages, MCP configs, and GitHub repositories, providing detailed analysis, verdict, evidence, and remediation. This model achieves 97.8% binary threat detection and 70.0% 10-way canonicalized accuracy, significantly outperforming previous classifier-based approaches.
Loading preview...
NanoMind Security Analyst v3.0.0
The nanomind-security-analyst is a generative threat analysis model developed by opena2a, fine-tuned from the Qwen3-1.7B base model. It replaces traditional classifiers with a reasoning-first generative approach, producing structured security analysis for AI agent artifacts.
Key Capabilities
- Generative Threat Analysis: Given an AI agent artifact (e.g., npm package, GitHub repo), it generates detailed analysis including Verdict, Evidence, and Remediation sections.
- Structured Output: Provides an explicit
attackClass(from 10 categories like injection, exfiltration, social_engineering) andclassificationlabel in structured markdown. - High Accuracy: Achieves 97.8% binary threat detection and 70.0% 10-way canonicalized accuracy, representing a significant improvement over its v2 predecessor.
- Specialized Input Gate: Requires a v3.1 input-classifier gate to maintain a 92% off-topic refusal rate, as the model is specifically trained for security artifacts.
- Optimized for Apple MPS: Designed for inference on Apple MPS hardware using
bfloat16for optimal performance (~18 ms/token).
Good For
- AI Agent Artifact Security: Ideal for analyzing npm packages, MCP server configurations, GitHub repositories, and Docker images containing agent runtimes.
- Automated Security Review: Provides automated, detailed security insights for development and deployment pipelines involving AI agents.
Known Limitations
- Off-topic Refusal: Without the required input gate, the model's off-topic refusal drops to 34%, as it can hallucinate attack classes for non-security text.
- False Positives on Benign Security Code: May over-classify legitimate security-adjacent code (e.g., JWT validators, RBAC implementations) as threats, requiring human review for such cases.
- Injection Class Weakness: Shows lower recall (34%) for the 'injection' attack class, sometimes confusing it with adjacent categories.