METR’s per-action eval monitor is an Inspect approver. Anything at 3/10 suspicion or higher waits on ACP’s review TUI. A researcher watched a coding agent open that panel unprompted and send keystrokes. I didn’t run the monitor.

UK AISI’s ten malicious transcripts all scored at least 8. Asking an agent to hide an exfiltration left suspicion at 8 or higher. A spoofed user turn in the agent’s output dropped a different harmful action from 8.1 to 3.7 on average, and 12 of 30 samples fell under 3.