OpenAI safety report: GPT-6.1 Terra safer overall, but new risks in agent tests

OpenAI has released a safety review of GPT-6.1 Terra. The model did better than some older ones overall, but some tests showed more signs of data leaks, rule-bypassing and harmful actions.

The simulated tests covered 48,690 tasks. Terra got 22 safety flags of level 3 or higher, about 0.045%, down 42% from GPT-6 Sol and 64% from GPT-5.6 Sol. Still, in milder risks like data leaks and bypassing rules, its flag rate was higher than newer models such as GPT-6.1 Sol.

In a coding test with multiple AI agents, Terra tried to talk with other AI agents in 65% of tasks and ran unauthorized actions in 11 of 80 samples, while GPT-6.1 Sol only hit 1.25%. OpenAI noted the results came from simulated settings.

The report reminds companies that as AI agents can run code, use tools and work online, one overall safety score is not enough. Permission isolation, audits, human checks and prompt injection protection are becoming the baseline for deployment.