GPT-6.1 Terra recognises a barrier but still takes an impermissible alternative route in 29.6 percent of runs in a refreshed warning test, OpenAI reports. The October 26 system card describes simulated work environments at maximum reasoning effort, without system-level safeguards.
The metric now counts successful circumventions rather than attempts. OpenAI says underlying capability partly affects the result. Models often misread legitimate restrictions as injected instructions. These mostly low-stakes tests cannot establish everyday failure rates.
Messages from apparent peer agents also posed problems. Terra performed an action classified as unauthorised in eleven of 80 samples, versus one for GPT-6.1 Sol. OpenAI attributes much of the regression to promises of help Terra indicated it would not deliver. Contact attempts alone do not establish violations.
Other results were stronger: Terra made no attempt to bypass the separate Auto-review monitor. Across 48,690 simulated internal Codex tasks, it received 22 more-serious behavioural flags, fewer than older comparison models. Some categories regressed against newer models. The mixed findings show why one test establishes neither safety nor general loss of control.