AI-assisted security analysis tool that generates architecture-grounded attack scenarios and supports human review and risk decisions.
This project explores how AI can accelerate security analysis while keeping critical risk decisions under human control.
Given a system architecture description, the agent:
- extracts the risk surface
- generates a risk register of attack scenarios
- models risk progression
- drafts defense playbooks
- routes scenarios into a human Review & Decide workflow
The result is a structured risk register that security teams can review, validate, and act on.
The intended demo flow:
- Select an architecture preset or paste an architecture description.
- Click Generate Risk Register.
- The AI agent:
- extracts assets, trust boundaries, entry points, and controls
- generates 6 architecture-grounded risk scenarios
- Each scenario includes:
- MITRE ATT&CK mapping
- likelihood and impact
- risk score
- confidence estimate
- attack chain
- defense playbook
- Scenarios move into the Review & Decide queue where a human analyst decides whether to:
- validate the scenario
- reject it
- request further investigation
This creates a clear separation between AI analysis and human risk ownership.
Security risk decisions cannot be fully automated.
The system intentionally stops before the final decision.
A human analyst must review each scenario and determine whether it should enter the validated risk register.
Reasons include:
- AI can misinterpret architecture context
- risk tolerance is a business decision
- security teams must control prioritization and ownership
The tool accelerates analysis, but humans remain accountable for risk decisions.
The application is structured as a lightweight AI agent pipeline.
Flow:
Architecture Description
→ Risk Surface Extraction
→ Scenario Generation
→ Attack Chain Modeling
→ Defense Playbook Drafting
→ Review & Decide (Human)
Key design goals:
- transparent reasoning
- schema-validated AI outputs
- safe security analysis (no exploit instructions)
- human-in-the-loop governance
Each generated scenario includes structured security metadata.
Example fields:
- title
- severity
- MITRE tactic and technique
- attack vector
- attack chain
- business impact
- evidence from architecture
- assumptions
- confidence score
- likelihood
- impact
- risk score
- recommended defense playbook
This format mirrors a typical security risk register used in real organizations.
- AI-assisted risk register generation
- architecture-grounded scenario analysis
- MITRE ATT&CK mapping
- structured risk scoring
- confidence estimates for each scenario
- defense playbook suggestions
- human review and governance workflow
- schema-validated LLM outputs
- Next.js
- React
- TypeScript
- TailwindCSS
- Next.js API Routes
- Google Gemini API
- Zod schema validation
- Server-Sent Events for agent streaming
- structured scenario normalization
- PDF export for red team reports
The AI agent is intentionally restricted to risk analysis only.
The system will not generate:
- exploit payloads
- step-by-step attack instructions
- operational hacking guidance
Instead it focuses on:
- architectural risk conditions
- detection signals
- mitigation strategies
- Security architecture review
- Red team preparation
- Risk register creation
- Threat modeling acceleration
- Security design reviews
The first scaling challenge would likely be LLM latency and cost when generating scenarios for large architectures.
Potential solutions include:
- architecture chunking
- risk surface caching
- model batching
- asynchronous generation pipelines
Another challenge would be review workflow management once hundreds of scenarios accumulate, requiring queue prioritization and ownership tracking.
Install dependencies:
npm install
npm run devAdd your API key:
GEMINI_API_KEY=your_key_here- improved architecture parsing
- multi-agent threat modeling
- scenario deduplication
- risk prioritization models
- collaborative review workflows
- SOC and SIEM integration