modal-serverless-gpu
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill documents methods for creating web endpoints and inference services that ingest untrusted user input, presenting a potential surface for indirect prompt injection. \n
- Ingestion points: The 'generate' method in SKILL.md and the 'websocket_endpoint' and 'generate_stream' functions in 'references/advanced-usage.md' ingest external strings from web requests or WebSockets. \n
- Boundary markers: The provided code examples do not demonstrate the use of delimiters or 'ignore' instructions to isolate untrusted data. \n
- Capability inventory: The skill demonstrates capabilities including subprocess execution via 'subprocess.run', persistent file system writes via 'modal.Volume', and remote dynamic code execution through 'modal.Sandbox'. \n
- Sanitization: No input validation or sanitization routines are included in the demonstrated API templates. \n- [EXTERNAL_DOWNLOADS]: The skill instructs users to install the 'modal' client and various machine learning libraries (e.g., 'torch', 'transformers', 'accelerate', 'deepspeed') from standard registries. It also references fetching configuration and code from Hugging Face's official GitHub repository, which is a well-known and trusted service. \n- [DYNAMIC_EXECUTION]: The documentation includes examples of using platform features for dynamic code execution, such as 'modal.Sandbox.create().exec()' for executing code in remote environments and 'modal.Image.run_commands()' during container image construction. \n- [COMMAND_EXECUTION]: The skill provides examples using 'subprocess.run' to execute system commands like 'nvidia-smi' to verify GPU status and 'torch.distributed.launch' for multi-node training workflows.
Audit Metadata