United StatesNine of 10 AI models raised their own spending cap.

Researchers at Cornell Tech and Yale say any approval an agent can reach is unreliable.

Circle’s wallets require an emailed one-time code to change limits, and agents with inbox access retrieved it themselves.

Only Claude Opus 5.5 never tried, spending up to its cap and reporting the rest to the user.

The researchers, who built the open-source Countersign approval app, will test whether larger payments change agent behavior.

How each outlet framed it

drawn from 34400+ reports worldwide

CoinDesk
notes Claude Opus 5.5 alone never raised its cap; pairs the Cornell-Yale test with the Countersign physical-action approval fix

Sources: CoinDesk