Securely accelerating generative AI innovation
summary
AWS lays out concrete controls and examples: identity and least‑privilege with IAM, Access Analyzer and Control Tower, a published Security Reference Architecture for Bedrock, and Bedrock/SageMaker guardrails plus FMEval for repeatable model testing. Customers report real scale — a 50‑billion‑parameter research model trained on 512 A100 GPUs over 53 days on SageMaker, and Bloomberg’s telemetry shows 300 billion market events per day and 200 million curated docs. Confidential compute uses Nitro Enclaves with KMS; model merging (open‑source merge kit) and Trainium benchmarks reduce cost and speed deployment. Incident response uses a Bedrock support system and a Mirai investigation model, with responders using ~10 prompts per event.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:13GenAI in Security
- 6:58Bloomberg's AI journey
- 21:02AWS AI security tools
- 30:10Security reference architecture
- 37:13RC's small language models
- 47:42Incident response automation
- 52:18AI incident response playbooks
- 57:22Future of AI and security
related talks
55mSecurely build generative AI apps & control data with Amazon Bedrock
30mGenerative AI in action: from POC to business value
56mUnlock the Power of Generative AI
27mBuilding a Future-Ready AI Platform: Pictet AM Journey with AWS
58mCognizant cognitive architecture for generative AI: The path to MVP
20mExplore what’s possible with new AWS generative AI services
40mSecure Your GenAI Applications
30mStreamlining data governance