The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requests broad permissions via allowed-tools to run pip install and python commands to facilitate SDK usage.
[REMOTE_CODE_EXECUTION]: Documentation provides examples using eval() and code_execution(True) capabilities. These patterns can allow arbitrary code execution if the agent processes untrusted inputs without sanitization.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading the inferencesh package and remote content for 'skills' via URLs.
[PROMPT_INJECTION]: The skill creates an architecture vulnerable to indirect prompt injection by allowing agents to ingest data from files, web search, and tool outputs while maintaining capabilities like code execution. Evidence: 1. Ingestion points: agent.send_message, web_search, upload_file. 2. Boundary markers: No delimiters or explicit warnings are suggested in the patterns. 3. Capability inventory: Access to code_execution, Bash, and network webhooks. 4. Sanitization: No input validation or filtering logic is included in the documentation.

python-sdk