data-engineering-streaming

Warn

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: MEDIUMPROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
  • Indirect Prompt Injection (MEDIUM): The skill demonstrates patterns for ingesting untrusted data from external platforms. \n
  • Ingestion points: Data enters the agent context via consumer.poll() (Kafka), on_message() (MQTT), and js.subscribe() (NATS) in SKILL.md. \n
  • Boundary markers: Absent; the code snippets do not include delimiters or instructions to ignore embedded commands within message payloads. \n
  • Capability inventory: The skill possesses network communication capabilities through stream client libraries and the requests library. \n
  • Sanitization: The code uses json.loads(), which validates structure but does not sanitize natural language content for potential prompt injection. \n- Data Exposure (LOW): The Kafka producer snippet uses socket.gethostname(), which exposes the local system's hostname to the message broker. \n- Network Operations (LOW): The MQTT example connects to broker.emqx.io, a non-whitelisted external domain. \n- External Downloads (LOW): The skill references the installation of legitimate packages (confluent-kafka, paho-mqtt, nats-py) from standard registries.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 16, 2026, 08:22 AM