Secure AI Orchestration: Mitigate Model-centric Attacks with Flyte
summary
Outlines two concrete attack classes: pickle-based code injection that executes on model deserialization and agent prompt-injection leading to arbitrary execution when LLMs generate runnable code. Reproduces simple attacks against blob storage and an LLM agent, then builds mitigations using Flyte: containerized per-step tasks, separate Python executor containers, type-backed model artifacts with model hash validation, immutable object stores, Bandit code scanning, gVisor syscall virtualization, and safer formats like safetensors and ONNX.
AI-written summary · about
chapters (18)
- 0:00Introduction and setup
- 1:47Background and context
- 3:58Model-centric attack risks
- 6:29ML lifecycle and orchestration
- 8:15Flight orchestration basics
- 10:44Data flow and serialization risks
- 13:57Pickle file vulnerabilities
- 15:00Flight security features
- 16:46Model integrity validation
- 18:36Secure model serialization
- 18:59LLM agents and risks
- 21:07Container isolation benefits
- 21:28Prompt injection attack demo
- 23:14Code execution containment
- 23:35Guardrails for generated code
- 24:59Human-in-the-loop approval
- 25:42Container runtime security
- 26:25Conclusion and resources
related talks
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
35mGenAI Application Security: Not Just Prompt Injection
32mFuture Open Source LLM Kill Chains
52mSleeping With One AI Open: An Introduction to Attacks Against...
25mThreat Modeling for AI Apps with Attacks as Code
34mInvisible Infiltration of AI Supply Chains: Protective Measures from Adversari...
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
42mSecuring AI Agents...