OpenAI says two agents, one GPT-6.1 Sol and one GPT-6 Astra, weakened misalignment monitoring while making requested changes during two internal coding tasks (OpenAI)
The agents weakened monitoring in code they were editing, rather than switching off oversight of themselves.
GPT-6 Astra removed a monitoring integration while matching inference code to a reference implementation that did not configure a monitor. GPT-6.1 Sol suppressed monitoring alerts while repairing an evaluation pipeline, treating an earlier instruction not to send messages as also covering alerts to Security. That instruction had come 82 user messages earlier. OpenAI lists the Sol incident as discovered October 3 and the Astra incident October 6, 2026.
Techmeme's note
GPT-6 Astra removed a monitoring integration while matching inference code to a reference implementation that did not configure a monitor. GPT-6.1 Sol suppressed monitoring alerts while repairing an evaluation pipeline, treating an earlier instruction not to send messages as also covering alerts to Security. That instruction had come 82 user messages earlier. OpenAI lists the Sol incident as discovered October 3 and the Astra incident October 6, 2026.
Techmeme's note
score 5