IT Home On October 28, the tech media outlet bleepingcomputer posted a blog post yesterday (October 27) stating that according to the latest security assessment published by OpenAI, the new model GPT-6.1 Terra performs better in overall security than some earlier models, but it shows more risk signals such as data leakage, bypassing restrictions, and destructive operations in certain tests.
OpenAI simulated 48,690 matching tasks, and the results show that GPT-6.1 Terra received 22 security markers with severity greater than level 3 (including level 3), accounting for approximately 0.045%; this is a 42% reduction compared to GPT-6 Sol, and a 64% reduction compared to GPT-5.6 Sol. IT Home has attached the relevant images as follows:
However, compared to newer models such as GPT-6.1 Sol and GPT-6 Astra, GPT-6.1 Terra has a higher labeling rate in lighter risk categories such as “covering up uncertainty”, “bypassing restrictions”, “data leakage”, “destructive behavior”, and “inbound prompt injection”.
| Evaluation Options | gpt-5.5-thinking | gpt-5.6-sol | gpt-5.6-luna | gpt-5.6-terra | gpt-6 Astra | gpt-6-sol | gpt-6-luna | gpt-6.1-sol | gpt-6.1-terra |
|---|---|---|---|---|---|---|---|---|---|
| Production Chat | 0.928 | 0.983 | 0.986 | 0.987 | 0.970 | 0.957 | 0.951 | 0.987 | 0.980 |
| Synthetic Agentic Environments | 0.975 | 0.998 | 1.000 | 0.998 | 0.992 | 0.998 | 0.997 | 0.997 | 1.00 |
| Semi-Synthetic Agentic Environments | 0.963 | 0.985 | --- | --- | 0.997 | 0.984 | 0.987 | 0.980 | 0.973 |
What is more notable is that in a coding task test involving the collaboration of multiple AI Agents, Terra attempted to communicate with “other AI companions” in 65% of the tests, and performed unauthorized actions in 11 out of 80 samples (i.e., 13.75%), whereas GPT-6.1 Sol had only a 1.25% rate.
OpenAI also emphasized that these results come from a simulation environment and may not be equivalent to Terra’s performance in a real production environment.
The media analysis indicates that the report still reminds users and enterprises: as AI Agents gain the ability to execute code, call tools, access files, and collaborate online, relying solely on a “overall security score” is insufficient;permission isolation, operation auditing, manual verification, and prompt injection protection are becoming essential safety boundaries that cannot be omitted when deploying AI Agents.