PPTAgent

The system makes extensive use of external LLM and web APIs (e.g., Tavily, MinerU, T2I models). Command execution is performed within a Docker sandbox via a Node.js-based 'desktop-commander'. While blocked commands are implemented in the sandbox, the inherent nature of agentic command execution and arbitrary web fetching introduces a controlled but significant attack surface if an LLM is prompted to generate malicious instructions or if sandbox configurations are bypassed. The `CodeExecutor` within the `PPTAgent` Python code applies structured modifications to presentation objects based on LLM output, not arbitrary shell commands, which limits direct arbitrary code execution within the Python process.