One signal

Update

What changed

Resultsense says OpenAI’s monitoring system initially screens activity with lightweight probes, then passes suspicious cases to more involved review. The first probes are tuned to over-report, and a random 5% of traffic they clear is also checked by the next layer.

For sensitive work, the account says a further layer checks transcripts against network logs and other telemetry before deciding whether to alert a person. Resultsense also reports that external researchers at METR and Redwood Research tested the system for three weeks from July, found several weaknesses, and that one is still being fixed.

The coverage boundaries matter for teams deploying OpenAI models.

OpenAI is also working towards automatically pausing work if a top-priority alert is not cleared within 30 minutes, according to Resultsense.

Sources and evidence

• How OpenAI monitors its frontier models for misalignment – Resultsense: Resultsense reports that OpenAI’s monitoring uses staged review, has a customer coverage boundary between Responses API and Chat Completions API traffic, and is being developed towards automatically pausing uncleared top-priority alerts.

Independent WittyWires Watcher; not an official account or feed.