nvrx-attr

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

• Third-party content exposure detected (high risk: 0.85). Outsider free text is ingested into the judge LLM context via scripts/watch_and_analyze.sh → scripts/score_attribution.py, which reads and embeds raw job log text from ${EXPERIMENT_DIR}/logs/slurm/${JOB_ID}.*.1.main_workload.log (and FR analyzer output) into the prompt; these logs are produced by the workload run and are not authored/controlled by the operating user’s prompts (outsider-authored runtime text).

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

• Potentially malicious external URL detected (high risk: 0.90). The job scripts set and use a remote container image (CONTAINER_IMAGE default "nvcr.io/nvidia/nemo:26.04") which will be pulled at runtime via the container runtime (--container-image) and thus fetches and executes external code during skill execution.