ODSP2185 Pre-recorded On demand
Secure Copilot with lessons from real-world AI vulnerabilities
Chen Levy Ben Aroy
Summary
From Microsoft's AI summary
Chen Levy Ben Aroy of Varonis Threat Labs described Copilot vulnerabilities the team found: Reprompt, using a 'queue' URL parameter; SearchLeak, a three-stage chain combining phishing and SSRF; and CoSnitch, a technique that surfaced an 'autorun=1' parameter. He argued that AI security equals data security. Recommendations included least-privilege access, treating AI assistants as high-risk identities, classifying sensitive data, limiting connectors and avoiding personal connectors such as Gmail or Google Drive. He also mentioned collaboration with Microsoft's Security Response Center.
Session description
Microsoft's catalogue description
As AI works across enterprise data, understanding emerging security risks is essential to using it with confidence. Explore three vulnerabilities identified by Varonis Threat Labs: Reprompt, SearchLeak, and CoSnitch, all responsibly disclosed and fixed in collaboration with Microsoft. Learn from this research and explore practical steps organizations can take to strengthen security as they adopt Microsoft Copilot.