MedDevice CyberGuard

AI-powered multi-agent system that produces FDA-ready cybersecurity risk dossiers for medical devices in minutes — using Amazon Nova to automate 6 weeks of manual compliance work.

Comment

Inspiration

In 2024, the FDA mandated cybersecurity documentation for all new medical device submissions under the Consolidated Appropriations Act. As a team with hands-on experience shipping medical device firmware under IEC 62304 in production, we have seen firsthand how manufacturers struggle with this requirement. Compliance teams spend 6 to 12 weeks manually preparing cybersecurity risk dossiers: mapping code vulnerabilities to FDA guidance, analyzing communication protocols, building threat models, and generating SBOMs.

Meanwhile, hospitals are deploying thousands of connected devices with unknown security postures. A compromised infusion pump or sterilizer is not just a data breach. It is a patient safety issue.

We saw an opportunity: what if AI agents could automate the entire cybersecurity assessment workflow and produce an FDA-ready dossier in minutes?

What it does

MedDevice CyberGuard is a multi-agent system powered by Amazon Nova that performs end-to-end cybersecurity assessment of connected medical devices.

Input: Device firmware (C/Python), IoT communication logs, network architecture diagrams, device specifications

Output: A structured cybersecurity risk dossier containing:

  • SAST findings with medical device context (patient safety impact, IEC 62304 safety class)
  • Protocol conformance analysis (IoT communication deviations from spec)
  • Regulatory compliance mapping (FDA, IEC 62443, IEC 62304, Health Canada)
  • STRIDE threat model
  • Software Bill of Materials (SBOM)
  • Architecture security analysis with IEC 62443 zone/conduit violations
  • Risk score and prioritized remediation roadmap

How we built it

Multi-Agent Architecture (Strands Agents + Amazon Nova 2 Lite):

  1. Code Security Analyst: Runs real Semgrep static analysis and Syft SBOM generation, then uses Nova to contextualize each finding for medical device risk (patient safety impact, attack vectors, exploitability)
  2. Protocol Conformance Analyst: Parses IoT communication logs and validates against the device specification, identifying deviations such as unauthorized clients, schema violations, forbidden topics, and QoS mismatches
  3. Regulatory Compliance Mapper: Uses RAG over FDA/IEC regulatory documents (Bedrock Knowledge Bases + S3) to map every finding to specific regulatory requirements
  4. Report Synthesizer: Uses Nova Multimodal Understanding to analyze architecture diagrams for zone segmentation issues, then merges all agent outputs into a unified dossier

Key Design Decision: Agents use real security tools (Semgrep, Syft, JSON Schema validators). Nova reasons over tool outputs rather than attempting raw code analysis. This ensures findings are accurate and defensible.

Infrastructure:

  • Amazon Bedrock (Nova 2 Lite for reasoning, Nova Multimodal for diagram analysis)
  • Bedrock Knowledge Bases with S3 data source (regulatory documents)
  • AWS Lambda (containerized agents with Semgrep/Syft)
  • DynamoDB (scan state), S3 (artifacts/reports)
  • React + Tailwind UI with real-time agent status

Cost optimization: Used S3 + Bedrock Knowledge Bases instead of OpenSearch Serverless, reducing RAG infrastructure cost from $700/month to ~$8 total for the hackathon.

Challenges we ran into

  1. LLM reliability for security findings: We initially considered having Nova perform SAST directly, but this produces hallucinated vulnerabilities. Solution: agents invoke real tools and Nova only interprets results.

  2. IoT protocol anomaly detection: LLMs cannot do statistical anomaly detection reliably. We reframed this as protocol conformance analysis: Nova compares logs against a specification document, which it excels at.

  3. Regulatory document copyright: We cannot distribute full IEC standards. We created comprehensive regulatory summaries that capture requirement structure and citations for RAG without violating copyright.

  4. Multimodal prompt engineering: Getting consistent structured output from Nova Multimodal for architecture diagrams required careful prompt design with explicit JSON schema and IEC 62443 zone/conduit terminology.

  5. Agent orchestration complexity: Coordinating 4 agents with inter-agent data dependencies. Strands Agents SDK made this manageable with clean tool definitions and structured output schemas.

Accomplishments that we're proud of

  • Real tools + AI reasoning: Not just "ChatGPT for security." Every finding is backed by actual static analysis, then enriched with medical device context.
  • Regulatory accuracy: RAG over real FDA/IEC guidance produces defensible citations that would hold up in a submission review.
  • End-to-end automation: From firmware upload to FDA-ready dossier in ~3 minutes. This genuinely solves a 6-week manual process.
  • Credible medical device expertise: The system understands IEC 62304 safety classes, IEC 62443 zone models, and FDA premarket requirements, not generic security advice.

What we learned

  • Multi-agent systems need real tools: LLMs are excellent at reasoning over structured data, poor at generating it from scratch. The best agent architectures combine classical tools with AI interpretation.
  • RAG quality depends on document preparation: Regulatory summaries with clear section structure and metadata outperformed raw PDF ingestion.
  • Bedrock Knowledge Bases are underrated: S3-backed RAG with zero ops overhead is ideal for early-stage products.
  • Medical device cybersecurity is massively underserved: Every manufacturer we have spoken to about this problem immediately wants it. The market need is real.

What's next for MedDevice CyberGuard

  1. Open-source the core: Publish the agent orchestration and medical device Semgrep rules so small manufacturers can self-assess
  2. CI/CD integration: Run assessments automatically on every firmware commit
  3. Hospital procurement integration: Enable hospitals to request CyberGuard dossiers as part of vendor due diligence
  4. Expand protocol support: Beyond current IoT protocols to cover CoAP, BLE, Zigbee, and cellular for broader connected device coverage
  5. Continuous monitoring: Webhook-triggered re-assessment and drift detection as device software evolves post-market
Share this project:

Updates