Detecting Security Posture Drift in AI-Driven Software...
summary
Prototype called trudge automatically derives threat models from code infrastructure and configs, computes threat model diffs across commits and releases, and flags five drift families: reachability, exposure, privilege, dependency, trust boundaries. It builds an intermediate canonical architecture graph for alignment and uses graph matching plus LLM reasoning. Evaluated on AWS: benchmark of 25 repos 219 commits finds one service at 97–98% precision/recall and another at 63% recall. Integration targets are developer inner loop, PR gating, or release-time checks.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:01AI agents in development
- 2:48AI-generated code risks
- 4:56Traditional SDLC gaps
- 6:43AI changes SDLC
- 8:49Threat modeling purpose
- 10:10AI security approaches limits
- 13:20Automated threat modeling
- 15:24Local static security problems
- 17:13Security drift in commits
- 19:57Intent versus deployed reality
- 20:59Continuous security tracking
- 22:46Canonical architecture tracking
- 24:10Computing security drifts
- 24:54Generalizing drift detection
- 25:55Security in AI workflows
- 26:16Key takeaways
- 26:59Q&A
related talks
40mState of (Absolute) AppSec
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
32mThreat Modelling in CI/CD environments
20mDynamic Defense: Crafting threat models with real-world Intelligence
43mDeveloper Driven Security In High-Growth Environments
14mAutomating Security and Compliance with AI-Based Analyzers
28mAI Under the Hood: Unmasking Hidden Threats
29mRedefining Threat Modeling: Security team goes on vacation