modal-serverless-gpu

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill provides examples for creating web endpoints and inference classes that ingest external text input (e.g., the prompt argument in TextGenerator.generate). This creates a vulnerability surface where malicious instructions embedded in the input data could be processed by the model or influence the agent's downstream actions.
  • Ingestion points: Found in SKILL.md (predict function and TextGenerator class) and references/advanced-usage.md (InferenceService and BatchInference classes).
  • Boundary markers: None provided in the examples to delimit untrusted input.
  • Capability inventory: The skill enables subprocess execution, file system access via Modal Volumes, and network operations.
  • Sanitization: No sanitization or validation logic is included in the provided code snippets.
  • [DYNAMIC_EXECUTION]: The skill documents the use of modal.Sandbox, which allows for the creation of isolated environments to execute arbitrary code and shell commands at runtime (e.g., sandbox.exec).
  • [COMMAND_EXECUTION]: Code examples utilize subprocess.run to execute system commands like nvidia-smi and to launch distributed training processes.
  • [EXTERNAL_DOWNLOADS]: The skill fetches base container images from NVIDIA's official registry and installs machine learning libraries from HuggingFace's public Git repositories.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 07:53 PM
Security Audit — agent-trust-hub — modal-serverless-gpu