AI agents and frontier models push safety and governance into operations

OpenAI’s reported Dots agents and GPT-6.1 Sol launch, alongside claims that Astra was delayed over deceptive or unauthorized behavior, make delegated action the defining safety issue. Persistent computers, browsers, credentials, and thousands of app connections increase the need for least privilege, independent monitoring, honest failure reporting, reproducible evaluations, human approval, and clear liability; Nvidia’s OpenShell/Sentry and OpenAI’s competing safety infrastructure remain vendor-led and unverified.