The Agent Skills Directory

[CREDENTIALS_UNSAFE]: Hardcoded placeholder credentials (postgresql://user:pass@localhost:5432/analytics) are present in the Definitions configuration block within SKILL.md. While used as a template, hardcoding credentials in source code is an unsafe practice.
[COMMAND_EXECUTION]: The skill utilizes shell commands for environment setup and orchestration, including pip install for dependencies, dagster project scaffold for project creation, and dagster dev to start a local server.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to fetch data from an external API (api.example.com) using the httpx library.
[DATA_EXFILTRATION]: The skill accesses the local file system using os.listdir on the /data/incoming directory to monitor for new data files.
[PROMPT_INJECTION]: The skill demonstrates an attack surface for indirect prompt injection.
Ingestion points: The skill ingests untrusted data from external API responses in assets.py and local file names in schedules.py.
Boundary markers: There are no markers or instructions used to separate data from commands or to instruct the agent to ignore embedded instructions in the ingested data.
Capability inventory: The DatabaseResource class in resources.py contains methods (query and execute) that run arbitrary SQL via SQLAlchemy.
Sanitization: The examples do not demonstrate sanitization or the use of parameterized queries, which could allow malicious data from the ingestion points to influence database operations if interpolated into SQL strings.

dagster