coder

The server implements several safety features, including automatic detection and handling of complex quoting for command execution via temporary script files, and an explicit 'DANGEROUS_COMMANDS' list for `run_command` that requires `require_confirmation=True`. It strongly advocates for specialized, safer tools (e.g., `write_file`, `read_file`, `search_code`) over raw `run_command` for file operations and searching, which significantly reduces the risk of shell injection. Output limits are in place to prevent memory exhaustion. There are no direct uses of `eval` or `exec` on unsanitized user input. The primary remaining risk is if an AI agent is instructed to use `run_command` with malicious shell code that circumvents the temporary script safeguards (e.g., by breaking out of the script context), or if `require_confirmation=True` is misused for destructive actions.