The Agent Skills Directory

[COMMAND_EXECUTION]: The script autodl_app.py executes the curl binary to interact with the AutoDL API and instance proxy services. It explicitly uses the -k (or --insecure) flag, which bypasses SSL/TLS certificate verification. This makes the transmission of sensitive data, such as the AUTODL_TOKEN, susceptible to man-in-the-middle (MITM) attacks.
[DYNAMIC_EXECUTION]: The script dynamically locates a sibling script, connect_server.py, by checking several potential paths within the user's home directory (e.g., .zcode, .codex, .cursor). It then executes this script using subprocess.run, which constitutes dynamic loading and execution from computed paths.
[DATA_EXFILTRATION]: The skill makes network requests to www.autodl.art and various dynamically discovered proxy domains under seetacloud.com. While essential for the skill's functionality, these operations involve sending authentication tokens and instance identifiers over the network to non-whitelisted domains.
[INDIRECT_PROMPT_INJECTION]: The skill is potentially vulnerable to indirect prompt injection because it processes JSON data from remote API endpoints and proxy services to determine the execution state (e.g., waiting for the instance to be ready).
Ingestion points: JSON responses from the AutoDL API and ComfyUI status endpoints on remote proxy domains.
Boundary markers: None. The script relies on direct JSON key access.
Capability inventory: Execution of shell commands (curl), local file access, and execution of Python scripts.
Sanitization: None. The script parses external JSON responses and uses the values directly for control logic without additional validation.

autodl-app-instance