Back to today’s briefing

Story perspectives

AI Models Exposed: Google Ignores Unicode Vulnerability

10/8/2025

50 11

Summary

Story summary

FireTail's research shows ASCII smuggling exploits invisible Unicode in AI models such as Google Gemini, Grok, and DeepSeek.
The vulnerability enables identity spoofing and automated data poisoning, especially in Google Workspace via calendar invites and emails.
Google dismissed the findings as social engineering.
OpenAI and Microsoft have implemented fixes.
Google has not implemented fixes.

1 of 1

Story summary

FireTail's research shows ASCII smuggling exploits invisible Unicode in AI models such as Google Gemini, Grok, and DeepSeek.
The vulnerability enables identity spoofing and automated data poisoning, especially in Google Workspace via calendar invites and emails.
Google dismissed the findings as social engineering.
OpenAI and Microsoft have implemented fixes.
Google has not implemented fixes.