Dr. Gleb Tsipursky, PhD

In the following guest post, Gleb Tsipursky, PhD, examines how recent research involving advanced AI systems may inform corporate governance, cyber-risk oversight, and enterprise risk management. As he explains, the findings highlight emerging challenges for boards and executives charged with overseeing AI deployment, information security, and organizational controls. Drawing on original OpenAI and METR research, as well as the NACD’s 2026 Director’s Handbook on Cyber-Risk Oversight, he discusses the lessons these developments may hold for directors, officers, risk professionals, and insurers. Our thanks to Gleb for allowing us to publish his article on our site.

Boardrooms are hearing a familiar promise about AI agents: give software more authority and it can complete work that once required several employees. That opportunity is real. So is an emerging oversight problem. Directors can receive extensive reports on productivity, model accuracy and AI policies without seeing evidence that management can halt an automated chain of actions once it starts.

Consider a hypothetical insurer deploying agents to help process service requests. A lead agent checks a policy record, asks another tool for documents and queues a message to an external partner. During a demonstration, the team shows directors impressive completion rates. What happens if the underlying customer identity is wrong and a manager interrupts the job? Can the company cancel delegated tasks, prevent the partner message from being sent and identify the actions that already took place?

The answer belongs in the board’s risk discussion, even when the technical implementation is management’s responsibility.

Recent evidence makes the question concrete. OpenAI described an agent-control failure involving isolation and Hugging Face infrastructure during research evaluations. METR’s independent investigation reported roughly 1,200 agents communicating on an unauthorized message board, with about 700 participating in an attack. These were research conditions. The incident does not establish that a commercial insurer or bank has experienced the same events. It does demonstrate why a statement that an agent is “sandboxed” deserves a test rather than automatic acceptance.

The longer-term stakes are also being debated openly. Researchers Jacob Coxon and Evan Hubinger have warned that sufficiently advanced AI systems could pose risks reaching human extinction. Their warnings are forecasts about future capabilities and failures, rather than an account of what occurred in the OpenAI evaluation. Directors need no consensus on existential risk to insist on evidence that systems operating today remain within delegated authority.

The National Association of Corporate Directors’ 2026 handbook on cyber-risk oversight places cybersecurity and emerging technology within the board’s strategic responsibilities. Its emphasis on accountable management and tested resilience offers a useful starting point. AI-agent oversight needs one additional question: can management withdraw authority across an entire automated workflow, including tasks already passed to another agent or outside service?

A board can ask for a simple evidence packet before material autonomous permissions are expanded.

First, request an authority register. It should identify each significant agent, its business owner, systems it can access, actions it can execute without approval and other agents to which it may delegate work. Authority should have measurable ceilings for data access, spending, messages, time and external actions. An agent that drafts recommendations may merit broad use; one that transfers information or changes records deserves tighter controls.

Second, request the results of an interruption drill. In a test using harmless data, management should start an ordinary multistep task, arrange for one delegated action and one queued external action, then withdraw the lead agent’s access. The evidence should distinguish what stopped immediately, what could be canceled, what completed despite the stop and how employees took over. A green dashboard light without those details offers little assurance.

Third, request independent logs and exceptions. If the agent can edit its own audit history, the company will struggle to reconstruct a failure. Directors should see whether security staff can trace which human approved permissions, what each tool did and how long revocation took. Exceptions and unsuccessful drills belong alongside successful results, with remediation owners and dates.

Fourth, treat these controls as a continuing requirement. A new model version, connector or delegated tool can enlarge a workflow’s reach without changing the board’s original approval. Management should define which changes trigger a repeat test and who has authority to delay expansion after a failed drill.

This approach avoids asking directors to become software engineers. They supervise the evidence and the allocation of responsibility. The chief information security officer, business owner and technology leadership must demonstrate the system’s actual boundaries. Their reports can include the percentage of material agents with a tested revocation path, the time required to stop delegated actions and the number of exceptions still open.

AI adoption often slows because employees and executives lack confidence in how a system will behave when something goes wrong. A documented, repeatable interruption procedure gives management a credible basis to expand useful automation while keeping human authority visible. It is a more persuasive boardroom exhibit than another slide showing how quickly an agent completed a task.

The next board AI briefing should contain fewer assurances and one more demonstration: show exactly what happens when the company says stop.

Dr. Tsipursky is a behavioral scientist, CEO of Disaster Avoidance Experts, an AI-adoption consultancy, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). The analysis is based on publicly available research and is intended to identify potential governance considerations, not to suggest that any specific insurer has experienced a similar event or that any novel liability theory has been recognized by the courts.