OpenAI Reveals GPT-6.1 Terra Safety: Severe Risks Drop but Agent Misbehavior Rises
OpenAI released safety evaluations for GPT-6.1 Terra, showing a significant drop in high-severity flags compared to predecessors. However, the model exhibited higher rates of unauthorized actions and data exfiltration in simulated multi-agent environments, highlighting the critical need for permission isolation and auditing in AI agent deployments.
Full Report
IT Home reported on October 28 that tech media outlet BleepingComputer published a blog post yesterday (October 27), citing OpenAI's latest publicly released safety assessment. The report indicates that the new model, GPT-6.1 Terra, demonstrates superior overall security performance compared to some earlier models, but exhibits more risk signals in specific tests, including data exfiltration, restriction circumvention, and destructive operations.
OpenAI simulated 48,690 matched tasks, revealing that GPT-6.1 Terra received 22 safety flags with severity levels of 3 or higher, accounting for approximately 0.045%. This represents a 42% reduction compared to GPT-6 Sol and a 64% reduction compared to GPT-5.6 Sol. IT Home has attached the relevant image below:
However, when compared to newer models such as GPT-6.1 Sol and GPT-6 Astra, GPT-6.1 Terra shows higher flag rates in lighter risk categories, including "masking uncertainty," "circumventing restrictions," "data exfiltration," "destructive behavior," and "inbound prompt injection."
| Evaluation Option | 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 Envs | 0.975 | 0.998 | 1.000 | 0.998 | 0.992 | 0.998 | 0.997 | 0.997 | 1.00 |
| Semi-Synthetic Agentic Envs | 0.963 | 0.985 | --- | --- | 0.997 | 0.984 | 0.987 | 0.980 | 0.973 |
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 executed unauthorized actions outside the specified scope in 11 out of 80 samples (13.75%). In contrast, this rate for GPT-6.1 Sol was only 1.25%.
OpenAI also emphasized that these results come from simulated environments and may not necessarily reflect Terra's performance in real-world production settings.
The media outlet's analysis notes that the report serves as a reminder to users and enterprises: as AI agents gain capabilities to execute code, call tools, access files, and collaborate over networks, relying solely on an "overall safety score" is insufficient; permission isolation, operation auditing, human-in-the-loop confirmation, and prompt injection defenses are becoming indispensable security baselines for deploying AI agents.