The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation files, specifically 'references/tool-builder.md', contain code examples that use the Python 'eval()' function to process input from tool arguments. This is an insecure design pattern that can be exploited for arbitrary code execution if a user prompt influences the tool's parameters.- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of agents with extensive capabilities, including code execution and web search, as documented in 'references/agent-patterns.md' and 'references/tool-builder.md'. These agents ingest untrusted data from user messages and tool outputs without providing explicit sanitization or boundary markers in the example templates. Ingestion points: Data enters the agent via 'agent.send_message()' and tool result submissions like web search outcomes. Boundary markers: The provided templates do not include delimiters or instructions to ignore instructions embedded in data. Capability inventory: The agents are configured with local code execution, network capabilities via webhooks, and web search access. Sanitization: The example code lacks validation or escaping for data processed at runtime.

python-sdk