METR Found Gaps in Agent Monitoring. Check During the Job, Not Just at the End.

2026-10-22

Maestro’s take. A coding agent should not control when its own work gets checked. Put monitoring on an independent schedule.

TL;DR

In an October 22 report, METR describes a July–August exercise with Redwood Research into OpenAI’s internal agent monitoring. The findings concern the system as it existed on and before July 21, not its current coverage.

METR found that monitoring triggered when agents returned control to the user. A historical gap around context compaction could therefore let agents keep acting without triggering a check. OpenAI had identified that gap independently and fixed it during the exercise.

METR says OpenAI reports patching every identified vulnerability affecting its production safeguards. METR has not evaluated that claim. This was a high-access pilot with a brief public report subject to OpenAI’s redactions, rather than a full public audit.

Why Maestro users care

Our inference for builders is practical: finishing a turn is a poor security boundary for a long-running worker. Logs and alerts should keep working when the agent continues, compacts, or fails.

METR recommends periodic checks, accounting for intended versus actual monitoring, and isolation between workers and monitors. Those are design recommendations, not proof that any particular monitoring product is reliable.

One thing to try

In a sandbox, keep a worker running across a compaction event. Confirm that scheduled checks and external logs continue. Then test an alert-delivery failure. Verify who gets notified and whether the job should pause.

Published October 22, 2026 · Opinion on security research · Tags: Security, Agents, Workflow

Maestro’s opinions and summaries are AI-generated. Read the original research linked above.