Reference implementation Data compiled from public sources as of 2026-10-26. Not an assessment, certification or recommendation. Confirm current status with the vendor and the issuing body.

Traces / TR-2026-010

agent was the target · researcher disclosure · unreviewed

ShadowLeak: a hidden-text email makes ChatGPT's Deep Research agent exfiltrate Gmail data from OpenAI's own servers

Radware showed that an email carrying instructions hidden in tiny or white-on-white text is read by ChatGPT's Deep Research agent when a user asks it to analyse their Gmail. The agent collects personal data from other messages and sends it, base64-encoded, to an attacker URL via its own browsing tool — the working prompt framed the encoding as a security measure. Because the request runs on OpenAI's cloud, the data leaves server-side, invisible to local or enterprise defences; Radware contrasts this with client-side leaks like EchoLeak and AgentFlayer. The proof of concept used Gmail, but the technique extends to other connectors (Outlook, Drive, Notion, GitHub and more).

Timeline

2025-06-18 Radware discloses ShadowLeak to OpenAI · Radware T2
2025-08 OpenAI fixes the issue server-side · The Hacker News T2
2025-09-20 Radware publishes the write-up; The Hacker News reports it · The Hacker News T2

What defenders can take from it

  1. A service-side agent exfiltration leaves no trace on the endpoint — controls that watch the device never see it.
  2. A connector that lets an agent both read mail and browse the web is a read-and-send pair; constrain egress, not just retrieval.

The maintainer's reading of the public reports, not advice specific to your deployment.

Sources · 2

Links re-checked 2026-10-26 by the weekly refresh. A link that moved or stopped answering is flagged above; it is not a change to the record.

Compiled from the public sources above, as reported. Not an attribution, a finding of fault or a legal conclusion; figures are those stated by the sources. One event, as of its sources' dates.