rocm-doctor
Pass
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: Downloads and executes installation scripts from the official ROCm GitHub repository (github.com/ROCm) to set up the necessary diagnostic tools. This occurs only when the CLI is missing and follows a consent-based workflow.
- [COMMAND_EXECUTION]: Invokes the
rocmCLI for system inspection and troubleshooting. The skill supports system-level fixes such as managing user groups or updating paths, which are gated by human verification and consent. - [PROMPT_INJECTION]: The skill ingests untrusted error data through a symptom-reporting flag. Ingestion points: User-provided error messages passed to the
rocm diagnosecommand. Boundary markers: Symptom text is encapsulated in double quotes within the command structure. Capability inventory: The skill can trigger system configuration changes (e.g., driver and permission fixes) via the CLI. Sanitization: Not explicitly detailed, but the CLI relies on a closed catalog of known symptoms to prevent arbitrary command execution via symptom text.
Audit Metadata