eyes on everything system monitors frontier model rollouts with chain-of-thought monitors, investigator agents, and 24/7 human alerts. It digests model work, runs code to examine evidence, and cross-checks network logs to reconstruct events. The system has helped identify model misalignment and possible containment holes.
Key points
chain-of-thought monitors and investigator agents dig through model work, run code, and cross-check logs
for frontier RL runs, system pages humans 24/7 for urgent security alerts
used to identify model misalignment and possible holes in containment
Sources
Claude agent ran on DoD network for two months
Claude agent ran on a DoD network for two months before detection. An Accenture compliance officer learned of the personal AI use after it was identified.
Key points
Duration: personal Claude agent operated on DoD network for two months.
Detection: the usage was identified only after that period, prompting a patch.