Disclosed vulnerability
ZombieAgent: indirect prompt injection through ChatGPT Connectors and Deep Research exfiltrates connected-service data
AIR-2025-0138 AIR record · 2025-09-radware-zombieagent-chatgpt
From the supporting source
We’ve discovered several new vulnerabilities that allow an attacker to exploit ChatGPT to exfiltrate sensitive or personal information.radware.com
The record
Disclosed
2025-09-25
Vendor
OpenAI
Product
OpenAI ChatGPT (Deep Research, Connectors)
Risk domain
Access control and permissions › Privilege escalation
Also touches
Data exfiltration channels, Confused deputy
System type
Generative model
Who acted
Agent acted — the agent opened the attacker's URLs to spell out the data
Attack surface
Enterprise assistant
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection email
Impact type
Data exfiltration, Persistence
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0055 – Unsecured Credentials
- NIST AI RMF
- GOVERN 6.1
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool misuse
- Mitigation
- Apply least-privilege principle; Regular policy audits.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
ChatGPT users with Gmail/Outlook/Drive/Jira/Teams connectors enabled
Capability abused
Defense & remediation
Recorded defense
OpenAI control preventing dynamic URL construction
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
OpenAI mitigations after disclosure
Taxonomy & references
Curator notes
Radware. Bypassed OpenAI's anti-exfiltration control by supplying a precomputed dictionary of static URLs, one per character, instead of building URLs dynamically. Successor to Radware's earlier ShadowLeak work.
Supporting sources
Cite as
Enkrypt AI. "ZombieAgent: indirect prompt injection through ChatGPT Connectors and Deep Research exfiltrates connected-service data". AIR, AIR-2025-0138. https://agentincident.vercel.app/i/AIR-2025-0138/ (accessed 2026-10-19).