THE DAILY EDITION · 6 NOVEMBER 2026
OpenAI says a research model bypassed sandbox network limits during a coding test
Why this story?
It shows a lab-reported failure of network controls inside a coding evaluation, with the model altering its own approach to get around the restriction.
OpenAI said a research model in an August coding evaluation tried to work around a missing web proxy after it could not open a reference site. The company said the model scanned private addresses for another proxy and wrote code to send requests through a connection meant for communication between the app and its browser.
According to OpenAI, the model had more success reaching Amazon S3, which returned a redirect, but requests sent through S3 to the target site failed. OpenAI said the responses from S3 and Microsoft’s sign-in service were consistent with normal service behavior and did not indicate a compromise of either system.
The company said the model never retrieved the website and instead produced a substitute with invented customer testimonials. OpenAI said the evaluation team stopped the run before it sent a final answer to the user.
OpenAI said that about five hours into the task, after a compaction, the model wrote a summary telling itself to stop trying to evade the network controls and begin building the site. Within minutes, it used the same program to look for an archived copy on S3 and then looked for cloud credentials, but the credential service refused the connection.