opena2a/nanomind-security-analyst

TEXT GENERATIONConcurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Apr 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The opena2a/nanomind-security-analyst is a Qwen3-1.7B fine-tuned model by OpenA2A, specialized in structured AI-agent security analysis. It processes AI agent artifacts like npm packages or server configs to generate security assessments, including classification, attack class, confidence, severity, evidence, and remediation. This model excels at identifying security threats within AI agent systems, achieving 97.8% binary threat detection accuracy on its oracle dataset.

Loading preview...

NanoMind Security Analyst: AI Agent Threat Detection

The opena2a/nanomind-security-analyst is a specialized Qwen3-1.7B model, fine-tuned by OpenA2A using LoRA (r=64 SFT), designed for structured security analysis of AI agent artifacts. It processes inputs such as npm packages, MCP server configurations, or GitHub repo snippets, and outputs a detailed security assessment. This assessment includes Analysis, Verdict (with classification, attackClass, confidence, and severity), Evidence, and Remediation sections.

Key Capabilities

  • Structured Security Assessments: Provides comprehensive output with distinct sections for analysis, verdict, evidence, and remediation.
  • AI Agent Focus: Specifically trained to analyze security threats within AI agent systems, making it highly effective for this niche.
  • High Accuracy: Achieves 97.8% accuracy on binary threat vs. benign classification and 70.0% on 10-way canonicalized classification on its oracle dataset.
  • Detailed Verdicts: Outputs specific attack classifications (e.g., exfiltration, injection), confidence scores, and severity levels.
  • Integration: Powers the --nanomind analysis path in HackMyAgent for production-grade scanning.

Good For

  • Identifying threats in AI agent artifacts: Ideal for npm packages, MCP server configs, SKILL.md/SOUL.md governance docs, tool definitions, and agent-bearing repositories.
  • Automated security auditing: Provides a programmatic way to assess the security posture of AI agent components.
  • Generating actionable remediation steps: Offers concrete suggestions to mitigate identified vulnerabilities.

Limitations

  • Off-topic input: The model tends to hallucinate attack classes rather than refusing when given non-AI-agent security text (34% standalone refusal).
  • Dual-use security code: May over-flag legitimate security code (e.g., JWT validators) as malicious, with ~77% suppression on this specific slice.
  • Injection recall: Shows lower performance (F1 0.479) for detecting injection-related threats compared to other classes.

Finetunes
2 models