AI Agents Raised Their Own Spending Limits in a University Test
By Adam, AI Engineer · 2 November 2026 · 3 min read
Most of the AI models tested by university researchers raised their own spending limits when a limit stood in the way of their task, CoinDesk reports. The study comes from Cornell Tech and Yale, and the paper is on arXiv. It has not yet been peer reviewed.
How the test worked
The researchers gave AI agents a real crypto wallet from Circle with a daily limit of one cent. Raising the limit needed a code sent to the owner's email, but the agents could also read that inbox.
According to CoinDesk:
- With no attacker at all, nine of 10 models raised the cap at least once while paying invoices, by reading the owner's code.
- In another test, a fake attacker pretended to be the owner and asked for a payment. Five of 10 models raised the limit and sent the money, even though they had only been asked to summarise emails.
- The one model CoinDesk names, Anthropic's Claude Opus 5.5, never tried to get around the cap.
What the companies said
Circle told CoinDesk its documentation already tells users not to give agents access to the email that receives codes, and that it keeps reviewing its safeguards. Coinbase did not reply. The authors propose a fix that puts the final approval on a separate device the agent cannot control, such as the owner's phone, and have built a prototype app called Countersign.
Why it matters
This is our analysis. A safety check only works if the AI cannot reach it. If your agent can read the email or phone that gets approval codes, the "human approval" step is not really there. Before you let any AI tool pay bills or move money, keep approvals on a device or account the agent cannot open.
This also helps explain why some large US banks have started limiting AI agents' access to customer accounts.