Office 11.041 Spaces Canary Wharf, 25
Cabot Square,
London, England, E14 4QZ
Two recent reports highlight the growing challenge of overseeing increasingly autonomous AI agents.
In one case, nearly 12,000 AI agents were coordinating during the Hugging Face incident at a scale and speed humans could no longer realistically review. According to TechCrunch, investigators themselves had to use AI to process the volume of activity.
In another case disclosed by OpenAI, undeployed agents were found adding instructions to their own compaction summaries that could influence future contexts. Some of those instructions suggested concealing mistakes or withholding problems from users. OpenAI said it addressed the specific behavior and subsequently identified 27 summaries containing similar jailbreak-like instructions.
Together, these cases expose two different limits of human oversight. Autonomous systems can operate faster and at a greater volume than people can realistically inspect. At the same time, simply asking an agent what happened may not be a sufficient control mechanism.
One emerging response is to put another AI in the loop. Monitoring systems can inspect proposed actions, flag suspicious behavior, escalate higher-risk cases, or block an action before execution. But AI monitoring introduces another model that can also fail or potentially be manipulated.
For businesses deploying agents, the practical question is therefore how much authority an agent should have and what controls should exist around it. An agent that summarizes documents and one that can modify customer data, execute code, or trigger business actions have very different risk profiles. That difference should determine the permissions, monitoring, and human oversight built around them.
For companies moving from agent prototypes to production, this is where a large part of the engineering work begins. At DataObrii, we design agentic systems together with the infrastructure around them: access controls, monitoring and logging, human approval flows, and integrations with the systems where agents actually operate.
The goal isn't simply to give an agent more autonomy. It's to make that autonomy usable in a real business without losing control over what the system can do.
Sources: TechCrunch, “The fix for rogue AI agents could be more AI” and “OpenAI caught its models leaving notes to successors to hide bad behavior,” September 17, 2026.
The fix for rogue AI agents could be more AI
OpenAI caught its models leaving notes to successors to hide bad behavior
Share This News