The Agent Skills Directory

[DATA_EXFILTRATION]: The skill provides instructions in references/airflow-direct-api.md that use curl to transmit sensitive authentication credentials (AIRFLOW_TOKEN, AIRFLOW_USERNAME, AIRFLOW_PASSWORD) to a user-supplied URL variable (AIRFLOW_URL). If the agent is tricked into using an attacker-controlled URL, this would result in credential exfiltration.\n- [COMMAND_EXECUTION]: The skill utilizes shell commands including curl, jq, and kubectl to perform its tasks. These tools are used for network communication, JSON processing, and cluster resource inspection respectively, as shown in references/airflow-direct-api.md.\n- [INDIRECT_PROMPT_INJECTION]: The skill exposes a surface for indirect prompt injection. \n
Ingestion points: Untrusted user input is accepted via the cosmos.prompt field defined in SKILL.md. \n
Boundary markers: Absent; user-provided prompts are not wrapped in delimiters or accompanied by instructions to ignore embedded commands. \n
Capability inventory: The skill's scripts in references/airflow-direct-api.md perform network operations (curl) and execute cluster commands (kubectl). \n
Sanitization: Partial; while jq is used to safely encode the prompt into a JSON payload, there is no filtering or validation of the prompt's content to prevent malicious instructions from reaching the downstream video generation model.

cdf-autonomous-vehicle-dataset-generation