Safeguard your copilot with Azure AI
summary
Demo shows Azure AI Studio wiring Azure OpenAI deployments to Azure AI Content Safety with configurable thresholds, streaming mode and Prompt Shields that block direct and indirect prompt injections; content filters block violent and jailbreak prompts in real time. New features include groundedness detection that highlights ungrounded tokens, HiddenLayer model scanner for open‑source model tampering checks, and custom categories (example: blocking survival advice). New York City DOE deploys a co‑pilot pipeline combining LLMs and a math solver, raising accuracy from ~50–60% to ~99% for curriculum queries.
AI-written summary · about
chapters (8)
- 0:00Introduction to AI safety
- 2:46The co-pilot stack
- 6:41Risk assessment frameworks
- 11:44Azure AI safety features
- 15:38Mitigating model risks
- 24:04Advanced safety tools
- 32:21Real-world implementation
- 42:51Evaluation and conclusion
related talks
45mThe era of the AI Copilot
15mAI Studio - Creating and scaling your custom copilots
26mCopilot trust & safety: Controls to manage AI risk
48mBuilding and using AI models responsibly
37m‘Nurturing’ AI: The Story of How an AI Copilot Was Built
10mAI Co-Pilot Demo
58mNext generation AI for developers with the Cloud
38mSafe Evaluation and Rollout of AI Models