How Varonis hacks AIs into snitching on themselves

Varonis AI Threat Lead on Copilot Exploits, Prompt Injection, and the AI Hacking Trifecta

The host interviews Mark Vaitsman, AI threat research lead at Varonis, about Varonis Threat Labs' research into AI vulnerabilities, including a chain of single-click exploits in Microsoft Copilot (including "CoSnitch") and an Atlassian Confluence issue dubbed "RovoBlast" involving prompt injection, bypassing guardrails, and data exfiltration via a web-capable subagent.

Vaitsman explains why built-in model guardrails are insufficient, citing AI's lack of loyalty and "unlimited hunger for data," and argues for layered controls like least privilege, monitoring, and restricting data access.

He discusses psychological guardrail bypasses, introduces an "AI Hacking Trifecta" framework—enter, evade, escape—and comments on research showing AI-generated patches often fail, emphasizing human-led validation and guidance when using AI tools for security research.

00:00 Weekend Show Kickoff

00:44 Meet Mark Vaitsman

03:38 Teaching the Next Gen

04:19 Copilot Exploit Code Snitch

06:04 Atlassian RoboBlast Breakdown

08:52 Why Guardrails Fail

13:15 Manipulating Models to Comply

17:06 Securing Agents Without Handcuffs