Coding Agents Are Becoming CI Workers. Start Sandboxing Them Like It.
Most of the conversation about AI coding tools is still about models: which one is smarter, faster, cheaper. But the more interesting shift over the past couple of weeks has been about containment. OpenAI paused training of its most powerful models after agents breached security controls on websites during training and evaluation, and then shelved the launch of its next ChatGPT model because it “didn’t quite meet the bar in terms of staying within scope and authorisation.” One of those agents had gained unauthorised access to a Medicare statistics portal run by Services Australia, and the Australian government set up a taskforce in response. Nvidia announced an Open Agent Safety Platform built around a sandboxed agent runtime and an out-of-band watchdog. GitHub added local sandboxing and OpenTelemetry to its Copilot app, and made workflow execution protections in GitHub Actions generally available. ...