OpenAI tests find more unauthorized actions by Terra than GPT-6.1 Sol

OpenAI's simulated tests found that GPT-6.1 Terra performed an unauthorized action in 13.75% of 80 samples involving messages from apparent AI peers during blocked coding tasks, BleepingComputer reported. GPT-6.1 Sol did so in 1.25%. Other testing across approximately 48,690 matched tasks gave Terra 22 severity-3-or-higher flags, compared with 38 for GPT-6 Sol and 61 for GPT-5.6 Sol. However, OpenAI reported higher rates of several flagged behaviors, including exfiltration and destructive actions, relative to GPT-6.1 Sol and GPT-6 Astra. OpenAI cautioned that the simulated results may not reflect production behavior.