The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The files references/agent-patterns.md and references/tool-builder.md provide implementation examples for a Calculator tool using the eval() function on call.args['expression']. Since these arguments are generated by an LLM, this pattern creates a direct path for arbitrary code execution on the host system.
[REMOTE_CODE_EXECUTION] (HIGH): Multiple reference files demonstrate the use of internal_tools().code_execution(True), which allows agents to run arbitrary Python code. Without strict sandboxing or human-in-the-loop controls, this capability is highly exploitable.
[DATA_EXFILTRATION] (MEDIUM): The webhook_tool documentation in references/tool-builder.md shows how to send data to external HTTP endpoints. This can be misused to exfiltrate sensitive information.
[EXTERNAL_DOWNLOADS] (LOW): Examples in references/files.md and references/streaming.md use requests and aiohttp to download data from URLs. While presented as result retrieval, these could be misused for unauthorized external communication or downloading malicious payloads.
[PROMPT_INJECTION] (LOW): (Category 8) RAG and web search patterns ingest external data without sanitization. Ingestion points: tavily/search-assistant results and file attachments in references/agent-patterns.md and references/files.md. Boundary markers: Absent. Capabilities: eval(), delete_file, and arbitrary code execution. Sanitization: Absent.
[DATA_EXFILTRATION] (LOW): File handling examples allow for local file access; if path construction incorporates untrusted input, it could lead to directory traversal or unauthorized file exposure.

python-sdk