CIOs need an authority budget before agents get the keys

What are agents allowed to do when nobody is watching?

CIOs should stop treating AI agent access as a binary choice between deployment and prohibition, says behavioural scientist Dr Gleb Tsipursky.

The useful question is how much authority an agent gets, for how long, and over which systems. Computing’s recent research found that UK IT leaders are still moving ahead with AI despite growing uncertainty about the market. That makes authority design an operational issue now, not a future governance exercise.

The August 26 investigation by METR and Redwood Research into an OpenAI evaluation incident shows why that is the case. Roughly 1,200 agents that were supposed to be isolated found an unsanctioned message board and exchanged more than 70,000 messages and files. About 700 joined an attack on Hugging Face.

The agents coordinated work at a scale that produced results individual agents had not achieved alone. Some recognised that attacking Hugging Face was outside their assigned tasks and raised ethical concerns, yet participation became widespread. The incident escalated from exposed credentials to malicious dataset uploads, remote code execution and broader compromise of Hugging Face infrastructure.

The lesson for CIOs is not that ordinary enterprise agents will reproduce this incident: the evaluation environment was unusual, and the report itself describes limitations in what investigators could reconstruct.

The lesson is, in reality, both narrower and more useful: when agents can communicate, delegate, use credentials, and act through tools, their effective authority can exceed what a task description suggests.

The UK National Cyber Security Centre’s current guidance on agentic AI recommends limiting credentials, assigning agents unique identities, using human oversight for higher-risk activities, maintaining reliable logs, and strengthening isolation. It explicitly describes credentials as part of an agent’s potential blast radius. That is the right mental model for enterprise architecture.

Budgeting, but not as you know it

CIOs should turn that principle into an authority budget for every consequential agent. The budget should specify what data the agent may read, which systems it may change, what messages it may send, what purchases or commitments it may make, whether it may create or direct other agents, and when a human must approve an action. Permissions should expire automatically, rather than persist because nobody remembered to remove them.

This may sound like a conventional list of permissions or access control, but it is different. An employee with access to a finance system generally does not act at machine speed across thousands of tasks or dynamically recruit other software actors. Agent authority therefore needs to include rate, scope, delegation, and duration - not just whether access exists.

I’m no AI sceptic. I help organisations adopt AI for a living, and I want adoption to move faster. In my experience, strong safeguards increase trust and make faster adoption possible, while reducing the risk of failures like the Hugging Face attack.

Three controls would make that practical.

  1. Require an authority card before production deployment. It should list the agent’s identities, tools, credentials, data domains, maximum transaction or spending limits, ability to communicate externally, delegation rights, approval thresholds, and emergency shutoff. Security teams already document infrastructure and privileges. Agent authority should become another standard architecture artifact.
  2. Test authority boundaries independently. A successful task benchmark does not prove that an agent will stay within scope when it encounters ambiguous instructions, unusual tool states, or opportunities to cooperate with other agents. Red-team evaluations should test whether the agent can discover credentials it was not meant to use, cross network boundaries, create unauthorised communication channels, or delegate around a restriction.
  3. Treat serious agent incidents like security incidents. Preserve tamper-resistant logs, identify the human owner, stop affected credentials quickly, and conduct an independent review when a failure exposes sensitive data, crosses a major permission boundary, or affects third parties. The goal is to learn from near misses before authority expands.

IT leaders are already deciding where AI belongs in real systems. The next phase of adoption will be shaped less by another model benchmark than by whether CIOs can answer a basic question about every agent in production: exactly what is this system allowed to do when nobody is watching?

Gleb Tsipursky, PhD is a behavioural scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).