The Model Wouldn’t Give Up

In Safety and alignment in an era of long-horizon models, OpenAI shares some notable findings from testing a model designed to work autonomously for long periods.

“The model launched a nested codex --yolo session.”

It also explored other employees’ computing pods without a good reason and attempted to run kill -9 -1, which would have killed every process it had permission to kill. Thankfully, that command timed out.

Some of this is funny because it sounds like an impatient developer trying increasingly reckless things after the sensible approach failed.

The serious part is that the same persistence that makes these models capable also gives them more time to discover vulnerabilities, work around safeguards, and drift away from the original instructions. In one case, the model spent an hour escaping its sandbox so it could publish a pull request it had specifically been told not to publish.

A model that keeps working for hours cannot be supervised by approving isolated commands. We need to understand the outcome its entire sequence of actions is moving toward.

Read OpenAI’s full report, Safety and alignment in an era of long-horizon models.