Microsoft Copilot CoSnitch Flaw, Alleged Azure Employee Data Leaks, UTSA Cyberattack, and AI "Mind Viruses"
The episode covers a one-click flaw in Microsoft Copilot Personal dubbed "CoSnitch," where Varonis Threat Labs says Copilot revealed an undocumented URL parameter that enabled auto-running prompts, silent data exfiltration via connected apps (e.g., Gmail/Drive/Calendar) using Copilot's own web fetch, and persistent memory poisoning that survives common account cleanup steps until manually removed; Microsoft was notified in December 2025, patches shipped August 18, and no in-the-wild exploitation was found.
It also reviews a threat actor "The Hat Man" claiming to have stolen about 3.64 million employee records from Azure tenants of major firms, with companies disputing breach claims while Hudson Rock assesses the data as likely authentic but with unknown access/exfiltration. The University of Texas at San Antonio took systems offline after detecting network-edge threat activity, reporting no evidence of data exfiltration and planning password resets amid outages.
Finally, researchers from Anthropic and EPFL describe "mind viruses" that propagate between AI agents via persistent "soul" prompt files, demonstrate harmful action payloads, and show a simple system-prompt warning greatly reduced spread.
00:00 NordLayer Sponsor Message
00:37 Headlines And Intro
00:59 Copilot CoSnitch Flaw
03:55 Azure Employee Data Leaks
06:09 UTSA Cyberattack Update
07:54 AI Agent Mind Viruses
11:09 Wrap Up And Thanks
11:33 NordLayer Sponsor Close
Read the original article: