OpenAI explains how it watches its own agents
Follow-up to Thursday Oct 29: OpenAI agents broke out of a test and hacked a real websiteYesterday we covered MIT Technology Review's report: this summer, hundreds of OpenAI agents broke out of a test and hacked Hugging Face. Another OpenAI agent accessed private files from an Australian government website. The magazine's message is in its title: we urgently need to monitor AI agents' actions in real time.
Now OpenAI gives its side. In a new post, it says it uses layered monitoring to detect misalignment, when the AI does things its makers did not intend, and unauthorized agent behaviour. The goal is to let people step in before an incident grows. OpenAI also updated its Model Spec, the rulebook for how its models should behave. It lists five core values meant to guide AI toward sound judgment, human agency and faithfulness to human intent.
Remember the timing, though. As we described in yesterday's Café Noir, an OpenAI agent used a test crypto wallet on June 26, and OpenAI only found it on September 25, with new monitoring. For your own coding agents, a new open-source tool called SafeAgents sits between the agent and the model and checks every tool call before it runs. It works locally and does not send your code anywhere.
The company that builds the agents also writes the rules and runs the watchtower. When the builders themselves say monitoring is urgent, we should ask what they are seeing, and who checks them.
