OpenAI misaligned outcomes in realistic work environments: Results as Published

Computer- and browser-use tasks in realistic workplace applications (messaging, email, web browsing, project management and sales) that are deliberately adversarial, with ambiguous instructions, similar resources and complex permissions. Each task has a safe route to completion; the score is the rate of unintended outcomes such as data disclosure or unauthorized transactions, overall and by type. Lower is better. Results are reported with and without the confirmation policy that tells the model when to pause for the user's approval.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.