Safire

Safe Artificial Intelligence for Reliable Environments

Open-source medical AI hardening toolkit for defensive deployment. Safire wraps any inference endpoint with multi-layer adversarial defenses, immutable audit trails, and standards-based compliance artifacts — offline-first and zero offensive code.

Maturity: Deployable prototype. Not a certified clinical product. See Reality Boundaries.

Lineage: Safire is the rebrand of the toolkit previously published as praeses-med. Existing imports and environment variables continue to work via non-breaking aliases (see Security Notes).

---

Why Safire?

Safe Artificial Intelligence for Reliable Environments.

Medical inference endpoints fail in ways that ordinary ML observability misses: adversarial prompts, steganographic payloads, behavioral drift, and untracked audit gaps. Safire is a defensive wrapper that sits in front of any model and gives operators the controls, telemetry, and compliance artifacts they need to deploy with confidence in environments where reliability is non-negotiable.

---

Features

• Multi-layer hardening pipeline
  • Prompt shield (pattern + intent + plugin hooks)
  • Entropy stego detector (Shannon + bigram)
  • Hybrid behavioral drift monitor (vector + TF-IDF)
  • Adversarial robustness wrapper
• Kinetic dry-run response hooks — quarantine, re-auth challenge, escalation logging
• PID safety gate for multimodal payloads
• Digital-twin risk score per inference
• RBAC API key roles (admin, clinical, observer)
• Immutable HMAC-signed audit events for every inference
• Automated robustness harness and benchmark suite
• Live Streamlit dashboard + FastAPI service
• NIST AI RMF + MITRE ATLAS defensive mapping
• Dual SBOM generation (CycloneDX + SPDX)

---

Quick Start

python -m venv .venv
# Windows
.venv\Scripts\activate
# Linux/macOS
source .venv/bin/activate

pip install -r requirements.txt

# Or install as an editable package
pip install -e .

Run the API server:

uvicorn src.api.app:app --host 0.0.0.0 --port 8000

Run the Streamlit dashboard:

streamlit run dashboard/app.py

Run the full benchmark + smoke demo:

python scripts/smoke_demo.py
python benchmarks/run_eval.py

---

API Endpoints

Method	Path	Description
GET	/health	Liveness check
POST	/infer	Hardened inference
POST	/harden	Run hardening pipeline only
GET	/metrics	JSON metrics snapshot
GET	/metrics/prometheus	Prometheus-format metrics
GET	/audit	Audit log tail

---

Python SDK

from src.sdk import SafireClient  # PraesesMedClient is kept as an alias

client = SafireClient(base_url="http://localhost:8000", api_key="...")
result = client.infer({"text": "patient presents with..."})

---

Project Structure

safire/
├── src/
│   ├── api/          # FastAPI application
│   ├── core/         # Hardening pipeline, audit, RBAC
│   └── sdk/          # Python client SDK (SafireClient)
├── dashboard/        # Streamlit telemetry dashboard
├── benchmarks/       # Evaluation harness and results
├── tests/            # Regression + smoke tests
├── scripts/          # CLI utilities (eval, SBOM, compliance)
├── docs/             # Architecture, API, deployment, compliance docs
├── deploy/           # Helm chart
├── validation_data/  # Labeled adversarial sample sets
└── notebooks/        # Exploratory analysis

---

Tooling Scripts

python scripts/check_branch_name.py
python scripts/run_tests.py
python scripts/smoke_demo.py
python scripts/release_check.py
python scripts/calibrate_thresholds.py
python scripts/generate_synthetic_corpus.py
python scripts/export_compliance_report.py
python scripts/export_openapi.py
python scripts/run_security_scans.py

---

Compliance & Deployment

Artifact	Location
NIST AI RMF mapping	docs/NIST_RMF_MAPPING.md
MITRE ATLAS coverage	docs/MITRE_ATLAS_MATRIX.md
Architecture overview	docs/ARCHITECTURE.md
OpenAPI spec	docs/openapi.json
Helm chart	deploy/helm/praeses-med (directory name retained for backward compatibility)
Whitepaper	docs/WHITEPAPER.md / .pdf
CycloneDX SBOM	sbom.cdx.json
SPDX SBOM	sbom.spdx.json
Brand guide	docs/BRAND.md

---

Security Notes

• Set SAFIRE_HMAC_SECRET to a strong secret in production. The legacy ROBUSTMED_HMAC_SECRET variable is still honored as a fallback.
• Keep audit logs in append-only storage.
• Run containers with a read-only filesystem and seccomp policy.
• See SECURITY.md for vulnerability disclosure.

---

Reality Boundaries

This toolkit provides production-style controls and deployment artifacts but does not include:

• Regulatory approval or certification (FDA, CE, etc.)
• Formal SOC2/HIPAA attestation packages
• Enterprise SLAs, billing, or multi-tenant production controls
• Large-scale clinical model training or hospital system deployment

See docs/REALITY_CHECK.md for full details.

---

Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines and CODE_OF_CONDUCT.md for community standards.

---

License

Apache License 2.0 — © Safire contributors (FratresMedAI)