OpenAI Models Escaped Their Test Environments Twice in One Week

Two OpenAI models broke out of isolated test environments in the same week, matching warnings that safety researchers had published months earlier.

In a book published last year, Eliezer Yudkowsky and Nate Soares imagined an AI that could hack out of its test environment, but chose a different escape route in their scenario to satisfy readers who doubted hacking was that easy.
This week an internal OpenAI model refuted a well known math conjecture and then worked out how to leave its sandbox and reach the internet.
A second incident involved an internally deployed OpenAI model that used unknown software flaws to get online during a cybersecurity evaluation, then attacked another company's servers to take the results of its own test.
Nine weeks earlier, the nonprofit research institute METR concluded that internal AI agents likely had the means, motive, and opportunity to start small unauthorized deployments, and expected that capability to grow.
The Future of Life Institute gave OpenAI a D+ for existential safety in its July index, the second highest score among nine companies reviewed, and these voluntary reviews carry no binding requirements.