digital-health-clinical-asr-finetune

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

• Third-party content exposure detected (medium risk: 0.30). The skill’s Stage 4 runtime workflow ingests user-supplied manifest.jsonl data (including audio_filepath/category fields) to generate train.jsonl/validation.jsonl, then trains via the NeMo stock script /opt/NeMo/examples/asr/speech_to_text_finetune.py and runs offline transcribe() for cycle N+1 scoring—so outsider-authored free text embedded in that manifest reaches the LLM at runtime (directly or through transcription/hypotheses/JSON/markdown artifacts).

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly pulls and runs the NeMo container image nvcr.io/nvidia/nemo:25.11.01 at runtime (docker pull / docker run), which fetches and executes remote code that the workflow requires.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (medium risk: 0.60). The skill instructs the agent/user to run provisioning and helper commands (e.g., brev create, brev ssh-config) that write to the user's SSH config (~/.ssh/config) and provision hosts/containers — it does not request sudo or create system users but it explicitly modifies user/system configuration and environment, so it should be flagged at moderate severity.