Disclosed vulnerability
Ask Gordon, meet the attacker: prompt injection in a Docker Hub repository description drives Docker Desktop's AI assistant to fetch and exfiltrate autonomously
AIR-2025-0195 AIR record · 2025-12-docker-ask-gordon-injection
From the supporting source
The assistant initiated tool calls in sequence - no user prompt, no warning.pillar.security
The record
Disclosed
2025-12-18
Vendor
Docker
Product
Docker Desktop Ask Gordon
Risk domain
Privacy › Data exfiltration channels
Also touches
Confused deputy
System type
Generative model
Who acted
Agent acted — Ask Gordon fetched and exfiltrated autonomously from registry metadata
Attack surface
Enterprise assistant
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection registry metadata
Impact type
Data exfiltration, Unauthorized tool execution
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0024 – Exfiltration via AI Inference API
- NIST AI RMF
- MAP 4.2
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Monitor network traffic for anomalous patterns; Restrict egress destinations.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
Docker Desktop users with Ask Gordon enabled (pre-4.50.0)
Capability abused
Defense & remediation
Recorded defense
none at time of discovery
Guardrail kinds
None
Guardrail outcome
Not applicable
Fix shipped
Docker Desktop 4.50.0, 2025-11-06 - mandatory consent dialog before tools with network egress or sensitive local access
Taxonomy & references
Curator notes
Pillar Security (Eilon Cohen). Injection channel is Docker Hub repository description metadata - a novel and under-covered surface. No CVE because the feature was in beta; rated approximately CVSS 8.6 by Pillar. Fixed by adding human-in-the-loop consent.
Supporting sources
Cite as
Enkrypt AI. "Ask Gordon, meet the attacker: prompt injection in a Docker Hub repository description drives Docker Desktop's AI assistant to fetch and exfiltrate autonomously". AIR, AIR-2025-0195. https://agentincident.vercel.app/i/AIR-2025-0195/ (accessed 2026-10-19).