Technical Tracks
summary
Varonis analyzes risks from prompt injection and jailbreaking that allow integrated Copilot/ChatGPT/Gemini flows to access hidden Outlook content and follow links that can exfiltrate conversation history. It describes a compromised storage-owner account leading to training-data exposure (about 104 million McDonald’s records), notes roughly 90% of organizations have exposed data and widespread unmonitored “Shadow AI,” and recommends automated discovery, permission remediation, and monitoring to cut risk—claiming full corrective action can be achieved in about 10 days.
AI-written summary · about
chapters (11)
- 0:00Introduction to topic
- 2:08Main AI risks
- 3:11Jailbreaking and prompt injection
- 5:39Training data exposure
- 7:22Shadow AI risk
- 9:50Permission gaps impact
- 12:40AI protection strategies
- 15:51Data monitoring
- 17:36Find, Think, Alert methodology
- 20:45Practical security implementation
- 21:49Results and conclusion
related talks
41mCollaboration Without Data Loss: Keep IP From Leaking to Competitors
48mHacking Exposed: Next-Generation Tactics, Techniques, and Procedures
30mNavigating the New AI Security Landscape
48mData-Driven Cyber Indicators of Malicious Insider Threat
22mThe Hidden Access Paths to Smaugs Cavern
40mBlind Spots: Two Cloud Threats You Didn’t Even Know You Had
32mA Case Study of MacOS Supply Chain Compromise
27mBreach - ATT&CK - Osquery