Protecting AI Infrastructure
summary
Claims that red teaming against DeepSeek yields 100% attack success on the top 50 HarmBench categories versus 26% for OpenAI, and that fine‑tuning raises jailbreak risk 3x and harmful responses 22x. Recommends machine‑scale validation, visibility and runtime guardrails, plus algorithmic red teaming. Announces an open‑weight 8 billion parameter model trained on 5B tokens (from a 900B corpus), designed for security reasoning, available on Hugging Face and able to run on one or two A100 GPUs for agentic SOC investigation and automated containment.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:49AI is fundamentally changing everything
- 3:05AI application architecture is completely different
- 6:19AI and model safety
- 8:39AI and validation
- 14:01AI and security
- 15:53Foundation AI
related talks
17mAI Safety in the Generative AI Era
16mKeynote: Defending AI on Kubernetes: Red Team Playbook
50mSecurity as Part of Responsible AI: At Home or At Odds?
27mKeynote: Towards Building Safe and Secure Agentic AI
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
19mKeynote: Towards Building Safe and Secure Agentic AI
20mAI Red Teaming Agent: Azure AI Foundry
50mThe ART of Probable: Test with AI, Atomic Red Team, and Threat Metrics