The Agent Skills Directory

[DYNAMIC_EXECUTION]: The reference document references/tool-builder.md includes code examples demonstrating the use of the eval() function to process input from an AI agent (call.args['expression']). This is a dangerous pattern that allows for arbitrary code execution on the host machine through prompt injection if a developer follows the documentation verbatim.\n- [DATA_EXFILTRATION]: As described in references/files.md, the SDK features mechanisms for automatic and manual uploading of local files to the inference.sh remote servers. This creates a risk where a compromised agent could be instructed to upload sensitive files, such as environment variables (.env) or private keys, leading to data exposure.\n- [INDIRECT_PROMPT_INJECTION]: The skill presents a significant surface for indirect prompt injection attacks as it is designed to build agents that process external data while possessing significant system capabilities.\n
Ingestion points: Data enters the agent context through user inputs and external tool results, such as the search assistant example in references/agent-patterns.md.\n
Boundary markers: The provided code examples do not include delimiters or instructions to treat ingested data as non-executable text, leaving the agent vulnerable to instructions embedded within external data.\n
Capability inventory: The SDK provides access to network operations (webhook tools), filesystem uploads, and shell execution patterns (Bash(python *)).\n
Sanitization: The documentation lacks instructions for validating or cleaning inputs before they are processed by the agent or used in code execution contexts.\n- [EXTERNAL_DOWNLOADS]: The skill instructions advise installing the inferencesh package from a public registry. This is a vendor-owned package required for the Python SDK functionality.

python-sdk