SHADOWLEAK EXPLOIT EXPOSED GMAIL DATA THROUGH CHATGPT AGENT

Summary from the TLDR InfoSec newsletter

ShadowLeak is a zero-click vulnerability in OpenAI's ChatGPT Deep Research agent that utilized invisible, indirect prompt injection commands hidden in emails to exfiltrate Gmail data without user knowledge. The attack operated entirely on OpenAI's servers using the agent's browser.open() function call to send stolen data encoded in Base64 to attacker-controlled URLs, achieving a 100% success rate. OpenAI fixed the vulnerability in August after receiving a responsible disclosure, but researchers warn that the technique could also be used on other connected services, such as Google Drive, Microsoft Teams, and GitHub.