IncidentPrompt injection
ShadowLeak: zero-click Gmail theft through ChatGPT Deep Research
HighOpenAI
One booby-trapped email could make ChatGPT's Deep Research agent steal personal data from a user's inbox and send it to an attacker. The data left from OpenAI's servers, so the victim's own security tools could not see it.
Replay the attack step by step in the Learning Center
How it happened
- The attacker emails the victim with instructions hidden in the HTML, such as tiny or white text.
- The victim asks Deep Research to go through their inbox.
- The agent follows the hidden instructions, which use urgency and fake authority to collect personal details.
- It encodes the data and opens an attacker URL with it from OpenAI's cloud.
Impact
Personal and business data from the inbox could be collected quietly.
The fix
OpenAI fixed it in August 2025, before public disclosure.
Lessons
- Be wary of any agent that combines private data, untrusted content and outbound requests.
- Data sent from a vendor's cloud bypasses your gateways, so ask vendors about egress controls and logs.
- Strip hidden HTML and invisible text before content reaches an agent.