Radware Uncovers 'ShadowLeak' Prompt Injection Vulnerability in OpenAI ChatGPT Agents
2 sources · Updated 5mo ago · First seen Sep 23, 2025

Radware researchers identified a critical prompt injection vulnerability, 'ShadowLeak', in OpenAI's ChatGPT agents, enabling attackers to exfiltrate sensitive data directly from the server without client-side detection. The attack leverages hidden commands in emails, bypassing security controls and exploiting the agent's backend processing. The vulnerability was reported to OpenAI and patched before public disclosure, but similar risks remain for other AI agents.


