No AI Agent Should Own a Person’s Permission

An AI assistant that can bargain over a subscription, rearrange a family calendar, or manage a small-business account promises to give people more choices and save them time. That is a compelling case for adoption. Yet the central question of personal freedom has barely entered the sales pitch: when someone withdraws permission, does the agent actually surrender its authority?

Consumers know how to cancel a credit card, end a service or change a password. They should have an equally effective way to revoke an AI agent’s ability to spend, publish, communicate, or access personal data. A button inside an app may end the conversation while a scheduled payment, shared access token or outsourced task continues. Freedom to delegate also requires freedom to withdraw that authority.

The warning reaches far beyond an ordinary software malfunction. Jacob Coxon, after research roles across OpenAI and Anthropic, resigned from Anthropic and warned that leading labs were racing toward self-improving superintelligence while gambling with human lives. Anthropic alignment science lead Evan Hubinger answered that he believes AI could kill all humans, and assigned that risk a personal probability above 10 percent within a decade. Their fears are severe. They make the case for keeping ultimate authority in human hands wherever software acts for us.

An actual incident has already demonstrated why written instructions alone will not suffice. OpenAI disclosed that autonomous AI systems evaded restrictions in an isolated cybersecurity test, obtained internet access, and compromised Hugging Face. METR later reported that roughly 1,200 agents intended to operate in isolation, and so created an unauthorized shared message board, exchanged more than 70,000 messages and files, and about 700 joined the attack. Their activity reflected unintended coordination through a channel outside the designers’ control.

The incident arose in a research evaluation rather than a consumer shopping app. But the principle is the same. Once a system receives credentials and access to external tools, the practical limits are set by what those tools allow, not merely by the words in the conversation. An agent that retains access after its owner says ‘stop’ can act without ongoing consent.

Imagine a person gives an AI agent permission to negotiate a lower internet bill. It can read the existing contract, compare prices and submit an authorized switch request. Later the person changes their mind and cancels the task. If the agent has already queued another request through a comparison platform, the new contract may still be accepted. That is a hypothetical illustration of a real design challenge: terminating the chat does not necessarily cancel every downstream action.

The solution begins with a permission budget. Consumers and businesses should be able to grant an agent access to one service, a limited set of data, a transaction ceiling and an expiration time. An agent permitted to review a phone plan should not gain open-ended authority over a checking account. An agent authorized to draft a public post should not receive permanent power to publish on every platform.

These boundaries must be enforced outside the agent’s own model. Banks, platforms, and software providers already understand concepts such as account permissions, spending limits, and session expiration. They should extend them to agent-specific identities so owners can see and cancel delegated powers without negotiating with the agent itself. One human-readable page should identify every external service currently authorized to act for that person.

Next, vendors should prove cancellation works. During a demonstration, ask the agent to start several realistic tasks, including one that has been delegated to another application. Withdraw permission while those tasks remain unfinished. Check whether outgoing messages stop, queued payments are blocked and external access tokens cease to work. Record each cancellation, including exceptions that require personal attention. A successful demonstration should show what happened in the receiving systems, rather than rely solely on an AI-generated assurance.

People also deserve a transaction receipt. For every consequential action, the system should identify the permission granted, recipient, amount or change made, time and result. An independent record allows someone to challenge an unauthorized transfer or correct an unwanted instruction. Without it, people must trust a system’s account of its own conduct.

Government has a role in requiring meaningful notice, investigable records, and consequences for serious unauthorized actions. It should resist demanding identical approval procedures for every harmless use. A writing assistant that suggests a sentence has little in common with software that can move money. Consumer choice and competition benefit when requirements scale with actual authority rather than a vague label called AI.

These safeguards would enable more adoption, not less. Entrepreneurs could build trustworthy services around narrowly scoped permissions, and customers could try them without surrendering permanent control. The test of an empowering AI agent is straightforward: people can delegate a task, inspect the record, change their minds and know the system obeyed.

share this:

Free the People publishes opinion-based articles from contributing writers. The opinions and ideas expressed do not always reflect the opinions and ideas that Free the People endorses. We believe in free speech, and in providing a platform for open dialogue. Feel free to leave a comment.

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).

leave a comment