内容
<p align="center">
<img src="https://img.shields.io/badge/version-0.1.0-blue?style=flat-square" alt="Version">
<img src="https://img.shields.io/badge/python-3.10+-3776AB?style=flat-square&logo=python&logoColor=white" alt="Python">
<img src="https://img.shields.io/badge/license-Proprietary-red?style=flat-square" alt="License">
<img src="https://img.shields.io/badge/status-Alpha-orange?style=flat-square" alt="Status">
</p>
# Eresus Sentinel
**Alpha-stage, deterministic-first AI security toolkit for local audits, MCP/agent checks, model artifact scanning, and prompt firewall testing.**
Sentinel provides deterministic, YAML-driven security scanning across the AI stack — from model artifacts and prompt firewalls to supply chain auditing and red team automation. Zero AI is required to produce findings; AI/judge adapters are optional enrichment layers.
<p align="center">
<img src="demo.gif" alt="Eresus Sentinel Demo" width="700">
</p>
---
## What It Detects
- **Serialization RCE** — Pickle/joblib/dill `GLOBAL` opcodes against 300+ blocklisted `module.function` pairs
- **Framework code injection** — Keras Lambda bytecode, TorchScript `code/` backdoor, PMML/Jinja2 SSTI, `auto_map` RCE
- **Archive exploits** — Zip Slip, path traversal, symlink attacks in ZIP / TAR / 7z / `.mar` / `.nemo`
- **Embedded secrets** — 2000+ API key patterns, connection strings, hardcoded tokens in model metadata
- **Network callbacks** — socket/urllib/requests globals in pickle; C2 URLs in GGUF metadata
- **Supply chain tampering** — missing hash integrity, typosquatting, OSV.dev CVE lookup
- **Prompt / agent injection** — instruction injection in model cards, MCP tool description hijacking
- **Backdoored weights** — statistical anomalies, entropy outliers, suspicious weight distributions (heuristic)
## Supported Formats
70+ scanners across model, archive, and configuration formats. **No model is ever loaded or executed.**
| Format | Extensions | Risk |
|--------|-----------|------|
| **Pickle / Joblib** | `.pkl` `.pickle` `.p` `.dill` `.dat` `.data` `.joblib` | CRITICAL |
| **PyTorch** | `.pt` `.pth` `.bin` `.ckpt` | CRITICAL |
| **TorchScript** | `.torchscript` (ZIP) | HIGH |
| **TorchServe** | `.mar` | HIGH |
| **Torch7 (Lua)** | `.t7` `.th` `.net` | HIGH |
| **NumPy** | `.npy` `.npz` | HIGH |
| **Keras** | `.keras` `.h5` `.hdf5` | HIGH |
| **TensorFlow** | `.pb` `.meta` `saved_model/` | HIGH |
| **Skops** | `.skops` | HIGH |
| **R Serialized** | `.rds` `.rda` `.rdata` | HIGH |
| **NeMo** | `.nemo` | HIGH |
| **CNTK** | `.dnn` `.cmf` | HIGH |
| **ONNX** | `.onnx` | MEDIUM |
| **JAX / Orbax / Flax** | `.jax` `.checkpoint` `.orbax` `.msgpack` | MEDIUM |
| **TFLite** | `.tflite` | MEDIUM |
| **ExecuTorch** | `.pte` `.ptl` | MEDIUM |
| **TensorRT** | `.engine` `.plan` `.trt` | MEDIUM |
| **CoreML** | `.mlmodel` `.mlpackage` | MEDIUM |
| **CatBoost / XGBoost / LightGBM** | `.cbm` `.bst` `.lgb` | MEDIUM |
| **PaddlePaddle** | `.pdmodel` `.pdiparams` | MEDIUM |
| **MXNet** | `*-symbol.json` `*.params` | MEDIUM |
| **SafeTensors** | `.safetensors` | LOW |
| **GGUF / GGML** | `.gguf` `.ggml` `.ggmf` `.ggjt` `.ggla` `.ggsa` | LOW |
| **OpenVINO / PMML** | `.xml` `.pmml` | LOW |
| **Llamafile** | `.llamafile` | LOW |
| **ZIP / TAR / 7z** | `.zip` `.tar` `.tar.gz` `.7z` | HIGH (traversal) |
| **RAR** | `.rar` | Fail-closed |
| **JSON / YAML manifests** | `config.json` `*.yaml` | HIGH |
| **Jinja2 templates** | `.j2` `.jinja` `.jinja2` | CRITICAL |
→ [Full format documentation with detection details](docs/SUPPORTED_FORMATS.md)
## Security Domains
| Domain | Module | Coverage |
|--------|--------|----------|
| 🔬 **Artifact** | `artifact/` | 70+ scanners — Pickle, Torch, Keras, ONNX, GGUF, Safetensors, TFLite, Archives |
| 🛡️ **Input Firewall** | `firewall/input/` | 22 guardrails — Injection, secrets, PII, encoding attacks, invisible text, toxicity |
| 🔒 **Output Firewall** | `firewall/output/` | 24 guardrails — Bias, compliance, copyright, watermark, format enforcement |
| 🔍 **SAST** | `sast/` | Static analysis + 120+ secret patterns + entropy + git history scanning |
| 🤖 **Agent/MCP** | `agent/` | Trust maps, permissions, MCP schema validation, live MCP discovery |
| 📦 **Supply Chain** | `supply_chain/` | Dependency scanning, typosquatting, OSV.dev, provenance |
| ⚔️ **Red Team / Eval** | `redteam/` | 48 probes + 13 detectors + 14 generators + YAML playbook/eval engine |
| 🔗 **MCP Proxy** | `mcp_proxy.py` | Live intercepting proxy (stdio/HTTP) with OPA policy enforcement |
| 🚦 **Runtime Gateway** | `runtime_gateway.py` | Provider-neutral LLM gateway contract with monitor/enforce modes |
| 📓 **Notebook** | `notebook_scanner/` | Jupyter security scanning (14 plugins) |
| 📝 **Diff** | `diff_scanner/` | Git diff/PR ML anti-pattern detection |
## Domain Maturity
| Domain | Maturity |
|--------|----------|
| Pickle/artifact no-load scanning | Beta |
| Prompt firewall deterministic checks | Beta |
| SAST/secrets/notebook/diff | Beta |
| MCP manifest/live scanning | Beta |
| MCP proxy runtime enforcement | Experimental |
| Dashboard/API | Experimental |
| HF/supply-chain live integrations | Experimental |
| Runtime gateway provider adapters | Experimental |
| AI/judge enrichment | Optional experimental |
## Quick Start
```bash
# Install
pip install -e ".[dev]"
# Verify installation
sentinel doctor
sentinel doctor --json
# Preview the scanner plan, then scan a project
sentinel scan ./my-project --plan --profile fast
sentinel scan ./my-project/
# Firewall a prompt
sentinel firewall "user input text"
# Scan model artifacts
sentinel artifact ./models/
# Red team assessment
sentinel redteam --target openai/gpt-4o
# Config-driven eval
sentinel evaluate eval.yaml
# Live/manifest MCP scan
sentinel mcp scan ./mcp-manifest.json
# Interactive shell
sentinel shell
```
## Architecture
```
┌──────────────────────────────────────────────────────────────┐
│ Eresus Sentinel CLI │
│ (Python · Click + Rich terminal) │
├──────────┬──────────┬──────────┬──────────┬──────────────────┤
│ Artifact │ Firewall │ SAST │ Agent/ │ Supply Chain │
│ Scanner │ (I/O) │ Analyzer │ MCP │ Auditor │
├──────────┴──────────┴──────────┴──────────┴──────────────────┤
│ YAML Rule Engine │
│ (rules/ — zero hardcoded regex) │
├──────────────────────────────────────────────────────────────┤
│ Finding Universal Data Model │
│ (7 domain factories + dedup fingerprints) │
├──────────────────────────────────────────────────────────────┤
│ MCP Proxy │ Eval Runner │ Runtime Gateway │ Telemetry │
├──────────────────────────────────────────────────────────────┤
│ AI-Assisted Layer (optional, pluggable) │
│ OpenAI / Anthropic / Local GGUF / Generic REST │
└──────────────────────────────────────────────────────────────┘
```
### Key Design Principles
- **Deterministic-first** — All scanning is regex/AST/opcode-based. No AI dependency for findings.
- **YAML-driven rules** — All patterns in `rules/*.yaml`. Zero hardcoded regex in code.
- **Plugin auto-discovery** — Drop a scanner class, it's automatically registered.
- **Lazy loading** — Modules load on-demand for fast CLI startup.
- **SARIF v2.1.0 output** — Native GitHub Security tab integration.
## CLI Commands
The most common commands are:
```bash
sentinel scan ./project --profile fast -f json
sentinel artifact ./models -f sarif
sentinel firewall "user input text" -f json
sentinel mcp scan ./mcp-manifest.json
git diff main...HEAD | sentinel diff - -f sarif
```
See [docs/CLI_REFERENCE.md](docs/CLI_REFERENCE.md) for the full command table
and [docs/CLI_CONTRACT.md](docs/CLI_CONTRACT.md) for exit codes and output
shape.
## Python SDK
```python
from sentinel import Sentinel
s = Sentinel()
# Scan input
result = s.scan_input("user prompt here")
print(result.findings)
# Scan output
result = s.scan_output("prompt", "llm response")
# Full conversation scan
result = s.scan_conversation("prompt", "response")
# Export to SARIF
s.export_sarif(result.findings, "report.sarif")
```
## Config-Driven Eval
```yaml
# eval.yaml
providers:
- id: local-echo
name: echo
prompts:
- id: greeting
prompt: "hello {{name}}"
tests:
- id: alice
vars: { name: Alice }
assertions:
- type: contains
expected: Alice
```
```bash
sentinel evaluate eval.yaml --fail-on-threshold 0.95
sentinel evaluate eval.yaml -f json -o eval-report.json
```
## MCP Live Scanner
Scan offline manifests or live MCP JSON-RPC endpoints:
```bash
sentinel mcp scan ./mcp-manifest.json
sentinel mcp scan --url http://localhost:3000/mcp
sentinel mcp scan --stdio-command npx my-mcp-server
```
The scanner discovers tools, prompts, resources, server instructions, auth metadata, and readiness signals.
## Runtime Gateway
```python
from sentinel.runtime_gateway import SentinelGateway, EchoProviderAdapter
gateway = SentinelGateway(provider=EchoProviderAdapter())
decision = gateway.complete("user prompt")
if decision.blocked:
print(decision.response.text)
```
## REST API
```bash
# Start Web UI dashboard
export SENTINEL_PASSWORD=change-me
sentinel dashboard
# Open http://127.0.0.1:8080
# If running from source and the UI is missing:
# Node.js 20.19+ is required by Vite/React Router.
cd frontend && npm install && npm run build
# Start API server
sentinel serve --host 0.0.0.0 --port 8080
# Scan endpoint
curl -X POST http://localhost:8080/scan/input \
-H "Content-Type: application/json" \
-d '{"text": "user prompt"}'
```
## MCP Proxy
Intercept and inspect all MCP protocol traffic in real-time:
```bash
# Stdio mode — wrap any MCP server
sentinel proxy --transport stdio --mode enforce --server-cmd npx my-mcp-server
# HTTP mode — reverse proxy
sentinel proxy --transport http --mode enforce --upstream http://localhost:3000 --port 8080
```
## Configuration
Engine is configured via `sentinel.toml`:
```toml
[engine]
mode = "deterministic" # "ai-assisted" | "full"
min_severity = "MEDIUM"
action_policy = "balanced" # "advisory" | "strict"
[scanners.artifact]
enabled = true
[scanners.firewall.input]
enabled = true
[scanners.firewall.output]
enabled = true
[scanners.sast]
enabled = true
[scanners.redteam]
enabled = false # Opt-in only
[ai]
enabled = false
backend = "ollama"
model = "llama3.2"
```
## Docker
```bash
# Standard
docker build -t eresus-sentinel .
docker run -v ./models:/data eresus-sentinel scan /data
# GPU-accelerated (CUDA)
docker build -f Dockerfile.cuda -t eresus-sentinel:cuda .
# Docker Compose (API + worker)
docker compose up
```
## CI/CD Integration
### GitHub Actions
```yaml
- name: Sentinel Security Scan
run: |
pip install eresus-sentinel
sentinel scan ./src --format sarif --output sentinel.sarif
- name: Upload SARIF
uses: github/codeql-action/upload-sarif@v3
with:
sarif_file: sentinel.sarif
```
## Documentation
**[Full Documentation](https://eresussec.com/docs)** | **[Quick Start](docs/QUICKSTART.md)** | **[Supported Formats](docs/SUPPORTED_FORMATS.md)** | **[Security Model](docs/SECURITY_MODEL.md)**
| Topic | Link |
|-------|------|
| Getting started | [Quick Start](docs/QUICKSTART.md) · [Turkish](docs/TR_QUICKSTART.md) |
| Format coverage | [Supported Formats](docs/SUPPORTED_FORMATS.md) |
| Scanner control | [Scanner Selection](docs/SCANNER_SELECTION.md) |
| Threat model | [Security Model](docs/SECURITY_MODEL.md) |
| Air-gapped deploy | [Offline Usage](docs/OFFLINE.md) |
| Python/OS compat | [Compatibility Matrix](docs/COMPATIBILITY.md) |
| CLI flags | [CLI Reference](docs/CLI_REFERENCE.md) · [CLI Contract](docs/CLI_CONTRACT.md) |
| Write rules | [Rule Authoring](docs/RULE_AUTHORING.md) · [YAML Rules](docs/RULES.md) |
| Write scanners | [Scanner Authoring](docs/SCANNER_AUTHORING.md) |
| CI / pre-commit | [CI and Pre-Commit](docs/CI_PRECOMMIT.md) |
| Reduce noise | [False Positive Handling](docs/FALSE_POSITIVES.md) |
| Ops | [MCP Proxy](docs/MCP_PROXY_DEPLOYMENT.md) · [Troubleshooting](docs/TROUBLESHOOTING.md) · [FAQ](docs/FAQ.md) |
## Authentication & Security
For production deployments:
```bash
# Enable API authentication
export SENTINEL_AUTH_TYPE=bearer
export SENTINEL_AUTH_TOKEN=your-secret-token
# Restrict CORS
export SENTINEL_CORS_ORIGINS=https://yourdomain.com
# Enable audit logging
export SENTINEL_AUDIT_LOG=/var/log/sentinel/audit.jsonl
# Rate limiting (configured via SENTINEL_RATE_LIMIT)
export SENTINEL_RATE_LIMIT=100/minute
```
See [Security Policy](SECURITY.md) for full hardening guide.
## Requirements
- Python 3.10+
- Core dependencies: `pyyaml`, `rich`
- API server: `pip install eresus-sentinel[api]` (adds `fastapi`, `uvicorn`)
- Web dashboard: `pip install eresus-sentinel[web]`
- ML scanning: `pip install eresus-sentinel[ml]` (adds `torch`, `transformers`)
- Rust pickle backend: `cd rust/sentinel-pickle && maturin develop --release`
- All extras: `pip install eresus-sentinel[all]`
## Alpha Disclaimer
> **This is alpha software.** APIs, CLI flags, finding schemas, and rule IDs may change between releases without deprecation notice. Do not depend on output stability in production CI pipelines yet.
## Community
- [Community Notes](docs/COMMUNITY.md) — contribution guidelines and communication channels
- [Good First Issues](docs/GOOD_FIRST_ISSUES.md) — beginner-friendly tasks
## License
Proprietary — © 2026 Eresus Security. See [LICENSE](LICENSE).
---
<p align="center">
<strong>Eresus Security</strong> · <a href="https://eresussec.com">eresussec.com</a>
</p>
连接信息
猜你喜欢
暂无推荐服务器