monitor-experiment

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

• Third-party content exposure detected (medium risk: 0.65). The required workflow fetches and ingests outsider-authored free text by SSH-ing into remote instances and running cat <results_dir>/<latest>.json / tail on screen hardcopy logs, where those files/logs are produced by external training jobs rather than the operating user’s own authored content.

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

• Attempt to modify system services in skill instructions detected (high risk: 0.70). The skill instructs the agent to SSH as root (e.g. "root@"), run remote commands that write files (hardcopy to /tmp), and suggests destroying cloud instances, which involve elevated privileges and changes to machine/cloud state.