Copilot trust & safety: Controls to manage AI risk
summary
Microsoft 365 Copilot embeds layered controls: proactive classifiers and sandboxing plus runtime guards that block prompt injection, hidden Unicode payloads, and unsafe outputs; demos show a sensitivity label that prevents summarization while still returning links, and an autolabeling flow that tags incoming mail. Admin tools include prompt DLP, restricted content discovery, protected‑materials classifiers, tool‑chain analysis, citation resilience and Safe Links; Microsoft ships 400 sensitive information types and allows domain exclusion of up to 250 domains for web grounding.
AI-written summary · about
chapters (10)
- 0:00Introduction and agenda
- 1:47Copilot architecture overview
- 3:01Proactive security controls
- 4:47Managing prompt injection
- 7:12Data discovery and web policy
- 9:41Runtime protection mechanisms
- 14:08DLP for Copilot demonstrations
- 17:43Auto-labeling configuration
- 18:47Egress and user controls
- 25:05Closing remarks
related talks
51mHow to Safely Deploy AI Copilots
47mSafeguard your copilot with Azure AI
13mPower Control and Security: Limitless with Microsoft Copilot Studio
11mCopilot: Bring protected AI chat to your organization at no additional cost.
47mData Security Considerations for AI Adoption
45mSecure and govern Copilot within enterprise-grade low code solutions
49mLiving off Microsoft Copilot
43mThe AI Assistant’s Betrayal: One-Click for AI to Turn into the Perfect Insider