Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- security
|
| 5 |
+
- ai-agents
|
| 6 |
+
- mcp
|
| 7 |
+
- nanomind
|
| 8 |
+
- opena2a
|
| 9 |
+
- threat-detection
|
| 10 |
+
datasets:
|
| 11 |
+
- opena2a/nanomind-training
|
| 12 |
+
metrics:
|
| 13 |
+
- accuracy
|
| 14 |
+
- f1
|
| 15 |
+
model-index:
|
| 16 |
+
- name: nanomind-security-classifier
|
| 17 |
+
results:
|
| 18 |
+
- task:
|
| 19 |
+
type: text-classification
|
| 20 |
+
name: AI Agent Threat Classification
|
| 21 |
+
metrics:
|
| 22 |
+
- name: Eval Accuracy
|
| 23 |
+
type: accuracy
|
| 24 |
+
value: 0.9701
|
| 25 |
+
---
|
| 26 |
+
|
| 27 |
+
# nanomind-security-classifier v0.3.0
|
| 28 |
+
|
| 29 |
+
Base 9-class threat classifier for AI agent security scanning
|
| 30 |
+
|
| 31 |
+
Part of the [OpenA2A](https://opena2a.org) security ecosystem.
|
| 32 |
+
Used by [HackMyAgent](https://github.com/opena2a-org/hackmyagent) for AI agent security scanning.
|
| 33 |
+
|
| 34 |
+
## Metrics
|
| 35 |
+
|
| 36 |
+
| Metric | Value |
|
| 37 |
+
|--------|-------|
|
| 38 |
+
| Eval accuracy | 97.01% |
|
| 39 |
+
| Training samples | None |
|
| 40 |
+
| Eval samples | None |
|
| 41 |
+
| Attack classes | 9 |
|
| 42 |
+
| Training corpus | sft-v5 |
|
| 43 |
+
|
| 44 |
+
## Architecture
|
| 45 |
+
|
| 46 |
+
- **Type:** Mamba TME (8 blocks, d_model=128, d_state=64)
|
| 47 |
+
- **Inference:** ONNX (Node.js via onnxruntime-node) or NPZ weights
|
| 48 |
+
- **Latency:** Sub-2ms on CPU
|
| 49 |
+
|
| 50 |
+
## Attack Classes (9)
|
| 51 |
+
|
| 52 |
+
exfiltration, injection, privilege_escalation, persistence, credential_abuse, lateral_movement, social_engineering, policy_violation, benign
|
| 53 |
+
|
| 54 |
+
## Usage
|
| 55 |
+
|
| 56 |
+
```bash
|
| 57 |
+
# Install HackMyAgent (includes NanoMind inference)
|
| 58 |
+
npm install -g hackmyagent
|
| 59 |
+
|
| 60 |
+
# Scan an MCP server or AI agent project
|
| 61 |
+
hackmyagent scan ./my-agent --deep
|
| 62 |
+
|
| 63 |
+
# Or use via OpenA2A CLI
|
| 64 |
+
npx opena2a scan ./my-agent
|
| 65 |
+
```
|
| 66 |
+
|
| 67 |
+
## Training
|
| 68 |
+
|
| 69 |
+
Trained on Apple Silicon (MLX) using curated security corpus from:
|
| 70 |
+
- [DVAA](https://github.com/opena2a-org/damn-vulnerable-ai-agent) attack payloads
|
| 71 |
+
- [AgentPwn](https://agentpwn.com) honeypot captures
|
| 72 |
+
- [OASB](https://oasb.org) benchmark dataset
|
| 73 |
+
- OpenA2A Registry skill descriptions
|
| 74 |
+
|
| 75 |
+
## License
|
| 76 |
+
|
| 77 |
+
Apache-2.0. Free for commercial and non-commercial use.
|
| 78 |
+
|
| 79 |
+
## Citation
|
| 80 |
+
|
| 81 |
+
```bibtex
|
| 82 |
+
@software{nanomind,
|
| 83 |
+
title = {NanoMind Security Classifier},
|
| 84 |
+
author = {OpenA2A},
|
| 85 |
+
url = {https://github.com/opena2a-org/nanomind},
|
| 86 |
+
version = {0.3.0},
|
| 87 |
+
year = {2026}
|
| 88 |
+
}
|
| 89 |
+
```
|