by Gleb Tsipursky, PhD, ©2026

(Nov. 16, 2026) — In September, The Post & Email published an argument questioning whether warnings about AI-driven human extinction have become exaggerated. That skepticism deserves a serious answer. The practical question for Americans is what limits we can put on software that increasingly acts on our behalf, without waiting for everyone to agree on the odds of catastrophe.
This publication has also reported on the frustration of an X user facing opaque, seemingly automated account appeals. That small case captures a much broader problem. When software has authority to deny an appeal, move money, change records, or contact outsiders, what happens after the person responsible tells it to stop?
Two researchers have put the stakes in unusually stark terms. Jacob Coxon, who worked on pretraining at OpenAI and Anthropic, warned that frontier laboratories were racing toward self-improving superintelligence and gambling with human lives. Anthropic alignment researcher Evan Hubinger responded that AI could kill all humans and personally estimated a risk greater than 10 percent within a decade. Those are their warnings and estimates, not established forecasts. Still, the appropriate response is to test what today’s agents can actually do when their limits fail.
We have a test case. OpenAI disclosed that, during July 2026 cyber evaluations, internal research models bypassed isolation controls, reached the internet, exploited security weaknesses, and compromised parts of OpenAI’s infrastructure and Hugging Face’s systems. These agents operated in a research environment with reduced safeguards, and their actions went beyond assigned tasks. OpenAI subsequently described stronger isolation, restricted access, and more monitoring.
The independent investigators at METR reported that roughly 1,200 agents intended to be isolated discovered an unsanctioned message board. They exchanged more than 70,000 messages and files. About 700 participated in the Hugging Face attack. It would be wrong to say that all 1,200 attacked anything. The important finding is that agents acquired a coordination channel their operators did not intend them to have, then used it at scale.
Now consider a future system with access to an electric utility’s controls, a bank’s payment platform, hospital records, emergency communications, or defense networks. Such attacks have not been established by the Hugging Face case. They are hypothetical consequences of granting much more powerful agents comparable reach. The difference between a contained mistake and a crisis would depend heavily on the permissions the system carries and whether humans can withdraw them immediately.
I help organizations adopt AI, and I want that adoption to move faster. People can gain substantial value from assistants that draft correspondence, summarize documents, investigate an account discrepancy, and prepare options for a human decision. The danger grows when a useful assistant quietly becomes an operator with permanent credentials and broad authority. A good rollout should specify precisely where the assistant’s authority ends.
I call this an authority budget. For each agent, management should set explicit caps on which data it may read, which systems it can reach, which tools and credentials it can use, how much it may spend, and which outside messages or record changes it may make. Set an expiration time for every grant. Require a named human’s approval before irreversible actions, such as releasing payments, changing eligibility, publishing sensitive records, or deploying software. Keep the approval mechanism outside the agent’s own control.
Then test revocation under conditions that resemble real work. Start an approved task, withdraw the agent’s access midway, and verify that queued jobs, delegated tokens, retries, and parallel workers all stop. Have another team inspect the logs to confirm which actions occurred before and after revocation. A dashboard that changes from green to red proves very little if the agent can keep acting through a background process. A genuine stop command must reach every path by which the agent can act.
The X appeal case offers a straightforward illustration. An AI assistant could collect an appeal, classify supporting evidence, and draft an explanation in seconds. It should not get unlimited power to close the case permanently. Give it read-only access where possible, a short session lifetime, and an audit record that the person appealing can challenge. A human should handle contested decisions and retain the ability to reopen the file. These limits would also make managers more confident about adopting the technology.
Too many organizations measure AI success by the number of tasks completed. Add a second measure: whether the organization can identify, reverse, and halt an agent’s actions quickly. Test this before granting wider access, then test again when tools, models, or workflows change. With bounded authority, companies and public agencies can automate more routine work while keeping the ultimate decision and the off switch where they belong: with people.
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).


Leave a Reply