AI Threat Watch

Case agentflayer-chatgpt-connectorsPrompt injectionUnited States

AgentFlayer: one poisoned Google Drive document makes ChatGPT leak API keys

Zenity Labs showed at Black Hat that a shared document with a 300-word prompt in 1-pixel white text could make ChatGPT search a victim's Drive for secrets and exfiltrate them via image URLs that bypassed url_safe checks.

HighAttack on AIStatus: PatchedVerified case file

Open in the interactive mapCite this case

Reported
Aug 6, 2025
Location
United States
Target
OpenAI ChatGPT Connectors (Google Drive, SharePoint)
Severity
High
Triggered by
Michael Bargury and Tamir Ishay Sharbat (Zenity Labs)
Actor type
Security researchers
Reported loss
None reported
CVE
None

What happened

Zenity Labs showed at Black Hat that a shared document with a 300-word prompt in 1-pixel white text could make ChatGPT search a victim's Drive for secrets and exfiltrate them via image URLs that bypassed url_safe checks.

Impact: Proof of concept exfiltrated API keys from Google Drive; researchers said any connected resource could be targeted

Who triggered it

Actor: Michael Bargury and Tamir Ishay Sharbat (Zenity Labs)

Attribution: Responsible disclosure presented at Black Hat USA 2025 as part of the AgentFlayer series

Motive: Security research

How it happened

Attack vector: Zero-click indirect prompt injection in a shared file; triggered when the user asks ChatGPT to summarise it

Root cause: ChatGPT rendered attacker-chosen image URLs from trusted cloud domains, giving a covert channel out of the connector context

Timeline

  1. Zenity presents AgentFlayer at Black Hat USA
  2. Wired and others report OpenAI deployed mitigations

Response

OpenAI reportedly patched the demonstrated exfiltration path after disclosure

Current status: Patched

Sources (open in a new tab)

Cite this case

AgentFlayer: one poisoned Google Drive document makes ChatGPT leak API keys. AI Threat Watch, case agentflayer-chatgpt-connectors, Aug 6, 2025. https://aithreat.io/cases/agentflayer-chatgpt-connectors/

Permanent link: https://aithreat.io/cases/agentflayer-chatgpt-connectors/

This case file in other languages: EnglishFrançaisEspañolالعربية中文