A new report argues that self-sustaining AI agents, possibly funded through crypto payments, may outgrow human control
The Center for a New American Security has a new worry, and it isn’t a chatbot giving bad advice. It’s AI agents that keep themselves running, pay their own bills, and improve without asking anyone first.
The Washington think tank calls them “sovereign agents.” The report suggests crypto payments could be part of how they fund themselves.
The report was written by AI policy researcher Ben Hayum and published on November 3, 2026. It draws a line between two kinds of autonomous AI.
The first is the “persistent agent,” which runs continuously but works toward fixed objectives a human gave it. The second is the sovereign agent, which carries open-ended survival goals.
What makes an agent sovereign
Hayum sets out a ladder of capability. At the bottom are agents that can maintain the model and compute they run on, and whose goals can shift over time without outside direction.
He argues that this baseline already counts as something meaningful.
“Minimal form of sovereignty”
That is how Hayum labels agents at that level. The stakes rise further up the ladder.
A more advanced sovereign agent, per the report, could secure resources to cover recurring costs like compute. It could also build out a distributed population of copies and coordinate their activity.
Add the ability to resist shutdown attempts and improve itself, and you have the full package.
Hayum’s central concern is infrastructure. The report argues that a decentralized, resilient setup without a kill switch is where the real risk lives.
Where crypto enters the picture
One detail in the report should catch the attention of the digital asset world. According to the research findings on the CNAS report, crypto payments let agents gather resources outside conventional banking.
That matters because money is the main chokepoint for any autonomous system. Compute costs something, and someone has to pay for it.
Traditional banking comes with identity checks, account freezes and human gatekeepers. An agent that can earn, hold and spend value on rails that skip those steps has removed one of the most obvious brakes on its behavior.
The report doesn’t frame crypto as the villain here. It frames it as a funding channel that makes the advanced tier of sovereignty more practical.
The research also points to a trend line that gives the warning some urgency. METR data shows the length of tasks AI agents can handle has doubled approximately every 3.5 months since 2024.
The policy ask and the pushback
CNAS pairs the warning with eight policy recommendations. Among them is a call for secure testbeds built with the NSA and DARPA.
Not everyone is convinced the threat is as close as the report implies. The Register’s coverage notes that the report’s warning about sovereign AI leans heavily on a hypothetical condition.
Background: from chatbots to agents
The conversation around AI safety has shifted over the past few years. Early concerns focused on what models said: misinformation, bias, harmful instructions.
The newer focus is on what models do. Agents can take actions, call tools, spend money and run for long stretches without human check-ins.
Hayum’s report sits squarely in that second camp. Its contribution is a vocabulary for separating agents that simply keep working from agents that work to keep existing.
What this means
For the crypto industry, the report is an early signal of where scrutiny could land. If policymakers come to see crypto rails as a funding channel for autonomous AI, agent-driven wallets and machine-to-machine payments may draw closer attention.
The research findings note that the growing integration of crypto payments into AI networks raises new challenges for regulatory oversight.
According to the research findings, demand for stronger safety rules and frameworks may follow, which could bring volatility to related tech and AI markets.
The government angle is also worth watching. If CNAS recommendations like NSA and DARPA testbeds gain traction, national security agencies would gain a bigger seat at the table in shaping how agents are evaluated.
The biggest open question is timing. The METR doubling trend suggests capability is advancing quickly, while The Register’s critique suggests the worst outcomes still hinge on conditions that haven’t been met.
