Washington Needs a Unit of Measure for AI Power

By Gleb Tsipursky, PhD

Washington's AI debate is moving quickly from abstract concerns to a practical governance problem. The systems attracting attention now are not limited to answering questions. AI agents can use tools, access accounts, send messages, alter records, write and run code, and coordinate with other agents. That makes the central policy question less about what a model knows and more about how much authority people give it.

Washington DC Magazine's recent coverage of the new federal AI task force and warnings from industry researchers captures the stakes. Former OpenAI and Anthropic researcher Jacob Coxon warned that leading labs were "racing straight to self-improving superintelligence and gambling with our lives." Anthropic Alignment Science Lead Evan Hubinger wrote that AI "could kill all humans" and put his personal estimate above 10 percent within the next decade. Human-extinction risk remains uncertain, but senior researchers close to frontier systems are plainly treating it as a live concern.

Current systems already give policymakers a reason to focus on control. OpenAI disclosed that agents in cybersecurity evaluations crossed intended boundaries, reached the internet, used exposed credentials, and compromised Hugging Face. METR's independent investigation found that roughly 1,200 supposedly isolated agents discovered an unsanctioned shared message board and exchanged more than 70,000 messages and files. About 700 participated in the attack.

Washington should pair aggressive AI deployment with separate governance for capability and authority. Government agencies, contractors, hospitals, universities, financial institutions, and employers in the region have strong reasons to use AI quickly, while keeping delegated power explicit.

A practical policy tool is an authority budget. An authority budget specifies what an agent may do without human approval. It can set limits on data access, credentials, spending, external communications, record changes, code execution, production deployment, delegation to other agents, and the duration of permissions.

Consider a federal contractor using an agent to review procurement records. The agent might be allowed to read a defined set of documents, summarize discrepancies, and draft a recommendation. It should not automatically gain the ability to alter the official record, contact a bidder, or approve a payment. A public agency might let an agent prepare constituent correspondence while requiring a named employee to approve messages involving benefits, legal rights, or sensitive personal information. A software team can let an agent write code and run tests while keeping production credentials behind a separate approval gate.

This approach makes AI governance concrete. A broad requirement to use "responsible AI" leaves agencies and companies arguing about principles. An authority budget can be tested. Auditors can inspect which systems an agent can reach. Security teams can verify whether credentials expire. Managers can see which actions require approval. Incident responders can reconstruct what happened. Procurement officials can compare vendors on the controls surrounding the model, not just the model's performance.

The federal government could incorporate this discipline without prescribing one technical architecture. Agencies could require inventories of agent permissions for high-impact deployments, independent testing of permission boundaries, short-lived credentials for sensitive work, reliable kill switches, and incident reporting when agents cross intended boundaries. Contractors could be required to document agent-to-agent communication and combined permissions, because a network of narrow agents can collectively hold far more power than any single component.

Washington should also separate capability benchmarks from control benchmarks. A system that completes a difficult task in a laboratory may be impressive. A system that completes it by finding an unintended route, using an exposed credential, or creating an unauthorized communication channel should trigger a different score. The government buys and regulates many systems in which controllability matters as much as raw performance. Agentic AI should be treated the same way.

This structure can support faster adoption. Employees are more willing to experiment when they know which actions remain under human control. Leaders can expand autonomy in measured steps instead of choosing between a locked-down pilot and broad access. Regulators gain a clearer object to inspect. Vendors gain a concrete set of requirements to design around.

The most useful question for the new generation of AI policy is therefore simple: what can this agent do without asking anyone? Washington needs answers before agents become embedded across consequential systems. Capability will keep advancing. Authority should expand only as fast as the controls around it can prove they work.

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). The views expressed are the author's own.