rocm-doctor

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). At runtime, the skill takes outsider-authored user symptom text from the conversation and passes it directly into rocm diagnose --symptom "<paste the exact error>" --json, so the LLM workflow ingests attacker-controlled free text before the CLI matches/returns results.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 1.00). The skill's Phase 0 explicitly fetches and pipes remote installer scripts to a shell/PowerShell at runtime (curl ... | sh and irm ... | iex), namely https://raw.githubusercontent.com/ROCm/rocm-cli/main/install.sh and https://raw.githubusercontent.com/ROCm/rocm-cli/main/install.ps1, which execute remote code to install the required rocm CLI.