Objection! AI Security Mistakes on Trial With Kubeflow and Confidential Computing
Talkat Cloud Native + Kubernetes AI Day Europe 2025 · on Tuesday, 2025-04-01· at 14:40· Level 1 | Hall Entrance N10 | Room G·
summary
Four recurring failures surface with concrete fixes and demos: untrusted model artifacts like pickle remote code execution that demand provenance checks and SBOM-style verifiers; generative errors where RAG grounding plus Kubeflow pipelines and OpenTelemetry tracing reduce hallucination rates; stolen credentials and API key theft from hardcoded secrets leading to platform jacking; and adversarial prompts where prompt injection / jailbreak is judged worst and is mitigated by system‑message hardening, NeMoGuardRails/Alam card classifiers, logging, and confidential containers with attestation.
AI-written summary · about
chapters (11)
- 0:00AI security mistakes
- 3:32Supply chain attacks in AI
- 6:42Hallucinations in generative AI
- 10:16AI platform jacking
- 13:06Prompt injection risks
- 15:54Worst AI security mistake
- 16:15Kubeflow for ML security
- 17:40OpenTelemetry for LLM monitoring
- 18:44Confidential containers
- 21:57Confidential containers demo
- 23:44Key takeaways
related talks
34mInvisible Infiltration of AI Supply Chains: Protective Measures from Adversari...
45mSecuring AI: There Is No Try, Only Do!
12mSecuring AI-Driven Finance
24mPanel Discussion: Proof in Production: Customer Lessons from Confidential AI
29mOwned By Statistics: How Kubeflow & MLOps Can Help Secure...
30mOwned by Statistics: Using Kubeflow to Defend vs Attacks on Your ML Models
53mHow to Break AI Systems (Before Someone Else Does)
57mAI Applications Security Puzzle