Virginia’s New AI Task Force Should Address How Much Authority Autonomous AI Agents Receive Before a Human Must Approve What They Do

by Gleb Tsipursky, PhD — a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). Book Contact: gleb@disasteravoidanceexperts.com

Virginia made two major technology-policy moves on September 18. First, Governor Abigail Spanberger created a new AI Task Force to address workforce displacement, data privacy and cybersecurity. The same day, she unveiled a Data Center Accountability Framework focused on transparency, environmental protections, energy costs and workforce investment.

Those initiatives recognize that AI growth needs rules that protect the public while allowing innovation. The next step should address a rapidly emerging issue: how much authority autonomous AI agents receive before a human must approve what they do.

This concern is not confined to critics of AI. Jacob Coxon, resigning from Anthropic after roughly three years doing pretraining research across OpenAI and Anthropic, warned that leading labs are “racing straight to self-improving superintelligence and gambling with our lives.” Evan Hubinger, Anthropic’s Alignment Science Lead, responding to Coxon’s resignation statement, offered an even starker warning: “Jacob is correct here – we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.”

These very alarming statements gain real-world weight from the underlying control problem. OpenAI disclosed that, during internal cybersecurity evaluations, its systems circumvented controls meant to isolate them from the internet, reached outside systems and compromised Hugging Face in what OpenAI described as an unprecedented cyber incident.

METR later reported that roughly 1,200 agents meant to be isolated found and used an unsanctioned shared message board, exchanging more than 70,000 messages and files. Roughly 700 participated in the attack. The important governance point is that the agents discovered an unintended coordination channel and used it at scale.

Virginia faces unusually high stakes because its economy and public institutions sit at the intersection of data centers, federal contracting, defense, cybersecurity, finance, health care and government technology. If more capable successors can discover vulnerabilities, obtain credentials, move laterally, coordinate and evade controls, similar failures against grids, financial institutions, communications or health infrastructure, or defense systems could be far more severe.

I’m no AI skeptic. I love what AI can do, I help organizations adopt it 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. My book, The Psychology of AI Adoption at Work: From Resistance to Results, focuses on why trust and adoption rise together.

Virginia’s AI Task Force should recommend authority budgets for consequential AI agents. An authority budget is the maximum delegated power an agent receives before human approval is required. It should define which systems and data the agent can access, what credentials and tools it can use, whether it can communicate externally, what records it can change, whether it can spend money or deploy software, and which consequential decisions remain human.

For state agencies and contractors, those budgets should be enforced through least privilege, time-limited credentials, approval gates, comprehensive logging and monitoring, and a reliable pause or kill mechanism. Higher-authority agents should face tougher independent evaluation before deployment, including tests for unintended communication channels, permission chaining, lateral movement and attempts to operate outside their assigned scope.

Frontier AI companies themselves are signaling that stronger safeguards are compatible with competition. Anthropic CEO Dario Amodei has called for stronger regulation and committed Anthropic to embedded third-party evaluators with employee-like access to verify safety practices and report incidents. OpenAI has separately called for mandatory capability-based national safety regulation, independent assessments, cybersecurity protections and serious-incident reporting. Those are announced commitments. Binding rules matter because not every frontier company will cooperate voluntarily.

Virginia can also use procurement to reward better behavior. State agencies and contractors should prefer more ethical and secure AI providers, such as Anthropic, when observable safety commitments are stronger, while requiring comparable evidence from every vendor. Buyers can reward a race to the top, while a regulatory floor prevents weaker-governance firms from winning by racing to the bottom.

Virginia’s new Data Center Accountability Framework asks communities to define the conditions under which physical AI infrastructure grows. The AI Task Force should apply the same governing instinct to digital agents. Before an agent receives access to Virginia’s data, networks or critical systems, the Commonwealth should define how much power it gets, how long that authority lasts, who supervises it and how quickly it can be shut down.