Security-Focused Chaos Engineering - the Lasso for AI Security T...
summary
An open source framework called Secops-Chaos encodes OWASP LLM risks and MITRE ATLAS TTPs as executable YAML experiments that run in Kubernetes and use verifiers to assert outcomes. Demos upload a Game of Thrones script with fake SSNs and AWS keys to the OpenAI API to detect input poisoning, then trigger prompt injection to show PII exfiltration, and finally deploy a third‑party pickle model in a privileged container to prove node data can be read when pod security/admission controls are missing.
AI-written summary · about
chapters (11)
- 0:00AI security and chaos intro
- 1:45AI threat modeling gaps
- 4:35Chaos engineering principles
- 7:28Microservices chaos examples
- 8:51LLM architecture risks
- 11:38Chaos for AI security
- 13:05SecOps Chaos tool
- 16:15Live demo: PII input validation
- 20:39Live demo: PII data leakage
- 23:35Live demo: third-party model risks
- 26:48Conclusion
related talks
25mThreat Modeling for AI Apps with Attacks as Code
40mSecurity & Chaos Engineering: Resilient Distributed Systems
28mBuilding Safety, Security and Resilience through Chaos
43mCyber Chaos Engineering
23mStandard Classification: Security Chaos Engineering - Turning the Tide in ...
40mNavigating the Unknowable: Resilience through Security Chaos Engineering
31mNavigating the Unknowable: Resilience through Security Chaos Engineering
42mDevSecOps for AI: Introducing MLSecOps for Software 2.0