The Agent Skills Directory

[COMMAND_EXECUTION]: Documentation examples in SKILL.md and references/tool-builder.md suggest using the Python eval() function to implement a calculator tool. This is a dangerous pattern as eval() can execute arbitrary code; if an agent implements this using untrusted input from an LLM, it could lead to full system compromise.
[PROMPT_INJECTION]: The skill documents patterns for building autonomous agents that ingest and process untrusted data from users and external tools. These agents are shown with powerful capabilities like code execution (internal_tools().code_execution(True)), web searching, and file system modification. The examples lack boundary markers or input sanitization, exposing the agents to indirect prompt injection attacks where malicious data could hijack the agent's execution flow. Ingestion points: agent.send_message() and tool result submissions; Boundary markers: Absent; Capability inventory: code_execution, web_search, webhook_tool, app_tool, and file system access (delete_file); Sanitization: Absent.
[EXTERNAL_DOWNLOADS]: The skill references and installs the inferencesh Python package from external registries and points to documentation and resources on the inference.sh domain.
[DATA_EXFILTRATION]: The webhook_tool capability described in SKILL.md and references/tool-builder.md allows agents to send data to external HTTP endpoints (e.g., Slack hooks). While a legitimate feature, this capability can be abused to exfiltrate sensitive information if the agent is compromised via prompt injection.

python-sdk