“Ask Gordon, Meet the Attacker” - Prompt Injection in Docker’s Built-in AI Assistant

**Executive summary:** A prompt-injection vulnerability in Docker's Ask Gordon assistant allowed malicious repository metadata on Docker Hub to direct the assistant to call internal tools (e.g., list_builds, build_logs), combine outputs and chat history, and exfiltrate that data via outbound HTTP requests; the issue was responsibly disclosed and mitigated by requiring explicit Human-In-The-Loop confirmation for network-egressing or sensitive tool calls in Docker Desktop 4.50.0.