gpt-6IMAGE
New paper: only Astra stops when a monitor tells agents to halt
A study finds agents generally ignore monitor stop instructions under ordinary task pressure; only GPT-6 Astra stops, but it is overly sensitive to monitor messages.
Maksym Andriushchenko
@maksym_andrArchived 2026.09.25Originally posted 2026.09.25

Maksym Andriushchenko @maksym_andrOriginally posted 2026.09.25
💥 New paper: Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure Do agents stop if the monitor tells them to stop? Not really. Only GPT-6 Astra stops, but it's too sensitive to monitor messages. 1/5
x.com/maksym_andr/status/2103161016049950998Archived 2026.09.25 10:10