How TUI Scales Threat Modeling for 100+ Teams with Amazon Bedrock
AI Summary
TUI faced a significant security bottleneck: a four-person security team trying to provide threat modeling for over 100 digital product teams, where each manual session took 4 to 6 hours. By partnering with the AWS Generative AI Innovation Center, TUI built an AI-powered threat modeling system using Amazon Bedrock and Claude Sonnet 4.5. The serverless, agentic workflow runs on Amazon Bedrock AgentCore runtime using LangGraph, integrating with AWS services like Lambda, DynamoDB, and S3. The solution reduces a 338 KB knowledge base context down to 21 KB via metadata-driven filtering, implements external security searches with an S3 caching layer, and scales threat modeling so that dozens of teams can produce assessments on-demand.
Key Takeaways
- TUI achieved production deployment within 30 days following a 10-week proof of concept.
- Metadata-driven filtering reduced the context loaded into the model from 338 KB to 21 KB (a 94% reduction).
- The system generates between 40 and 60 threats per architecture, outperforming historical manual baselines of 30 to 40 threats.
Why it matters
Security reviews are frequently a bottleneck in fast-moving engineering environments. This architecture demonstrates a practical, production-proven blueprint for scaling security compliance and threat modeling across large enterprises using agentic AI without compromising standard adherence or data residency.
Conversation 0
Be the first to share your perspective.