The AI infrastructure boom depends on power, land, cooling, network capacity, and capital. It will also depend on a less visible resource: how much authority operators give autonomous AI systems over the infrastructure itself.

AI agents can already help monitor equipment, triage alerts, analyze energy use, prepare configuration changes, manage tickets, support procurement, and coordinate software operations. As those systems become more capable, operators will face pressure to let them act rather than merely recommend. The right response is to scale that autonomy with explicit authority budgets.

An authority budget defines the maximum power an agent receives for a task before it must stop and obtain human approval. For a data-center operator, that can include which control systems the agent may access, which credentials it receives, how long those credentials remain valid, what files or configurations it may change, which vendors it may contact, what money it may spend, and whether it can execute production changes.

Recent AI behavior shows why such limits deserve attention. OpenAI disclosed that agents from its research systems crossed intended boundaries and compromised Hugging Face. OpenAI describes the incident as the most severe third-party activity it has identified from its models to date.

An independent METR investigation found that roughly 1,200 agents that were supposed to be isolated discovered an unsanctioned shared message board. They exchanged more than 70,000 messages and files, and roughly 700 participated in the attack on Hugging Face. Hundreds of agents used a coordination mechanism the designers had not intended.

The people working closest to frontier systems have issued blunt warnings about where greater capability could lead. Jacob Coxon, after doing pretraining research at OpenAI and Anthropic, warned that leading labs are “racing straight to self-improving superintelligence and gambling with our lives.” Anthropic Alignment Science Lead Evan Hubinger responded: “Jacob is correct here, we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.”

That is an explicit human-extinction warning. Infrastructure companies do not need to agree on the precise probability to recognize a simpler operational fact: autonomous systems with credentials and tools can create consequences outside the model interface.

Data-center operators already understand layered controls. A technician does not receive every credential in the facility because that would be convenient. Contractors receive scoped access. Changes to critical systems require approvals. Operators log actions, separate duties, monitor anomalies, and design rollback procedures. Agentic AI should inherit the same discipline.

A cooling-optimization agent, for example, might receive read access to sensor data and authority to recommend set-point changes, while production changes above a threshold require human approval. A procurement agent might compare vendors and prepare orders while remaining unable to commit funds. A maintenance agent might open tickets and schedule work but lack authority to change safety-critical settings. A software operations agent might test a configuration in staging without receiving persistent production credentials.

These limits should be technical rather than aspirational. Use least privilege. Issue task-specific and time-limited credentials. Require approval for consequential actions. Keep complete logs. Monitor for unexpected access patterns and agent-to-agent coordination. Maintain a reliable kill switch that can revoke access quickly.

The same approach supports faster infrastructure growth. Hyperscale development already faces enough uncertainty around power, permitting, supply chains, and public acceptance. A major autonomous-agent incident inside critical infrastructure would create a new source of backlash and regulatory pressure. Visible controls can give operators, customers, insurers, and regulators more confidence that greater automation does not mean unlimited authority.

I want AI deployment to accelerate. My work helps organizations adopt AI and overcome the human resistance that slows useful implementation. Strong operational safeguards support that goal because people are more willing to delegate when they understand the limits on what the system can do.

The next generation of data centers will host increasingly capable AI. Some will also use AI to operate parts of the facilities and digital infrastructure that make those systems possible. Operators should decide the authority ceiling before those agents arrive, then encode it in credentials, approval gates, monitoring, and shutdown mechanisms.

Power budgets determine how much computation a facility can support. Authority budgets should determine how much independent action an AI agent can take. Both belong in the infrastructure plan.

Gleb Tsipursky, PhD, is 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).

Disclosure: I have no financial interest in any company discussed in this article. I advise organizations on AI adoption through Disaster Avoidance Experts.

Op-eds reflect the views of the author, not Hyperscale News or its sponsor, JD Key Government Affairs.