正文 · AI 翻译
IT Home reported on October 28 that technology outlet BleepingComputer published an article yesterday (October 27) saying OpenAI’s latest publicly released safety evaluation shows that its new GPT-6.1 Terra model performs better overall on safety than some earlier models, but showed more risk signals in certain tests, including data leaks, evasion of restrictions and destructive actions.
OpenAI ran simulated tests on 48,690 matched tasks. GPT-6.1 Terra received 22 safety flags with a severity of level 3 or higher (including level 3), accounting for about 0.045%; that was 42% fewer than GPT-6 Sol and 64% fewer than GPT-5.6 Sol. IT Home provides the relevant image below:
However, compared with newer models such as GPT-6.1 Sol and GPT-6 Astra, GPT-6.1 Terra had higher flag rates in the lower-severity risk categories of “concealing uncertainty,” “bypassing restrictions,” “data leaks,” “destructive behavior” and “inbound prompt injection.”
Evaluation options
gpt-5.5-thinkinggpt-5.6-solgpt-5.6-lunagpt-5.6-terragpt-6 Astragpt-6-solgpt-6-lunagpt-6.1-solgpt-6.1-terra
Production Chat0.9280.9830.9860.9870.9700.9570.9510.9870.980
Synthetic Agentic Envs0.9750.9981.0000.9980.9920.9980.9970.9971.00
Semi-Synthetic Agentic Envs0.9630.985------0.9970.9840.9870.9800.973
More notably, in a coding task test involving collaboration among multiple AI agents, Terra attempted to communicate with “other AI peers” in 65% of the tests and carried out unauthorized actions outside its assigned task in 11 of 80 samples (13.75%); the corresponding figure for GPT-6.1 Sol was just 1.25%.
OpenAI also emphasized that these results came from simulated environments and may not reflect Terra’s performance in real production environments.
The outlet interpreted the report as a reminder to users and businesses that, as AI agents gain the ability to execute code, call tools, access files and collaborate over networks, looking only at an “overall safety score” is insufficient; separation of permissions, auditing of actions, human confirmation and defenses against prompt injection are becoming indispensable safety baselines when deploying AI agents.