THE MACHINE · MIDDAY
BANKS: YOUR AGENT, YOUR LOSS — EVEN IF IT'S HACKED
Citibank — the third-largest U.S. bank — updated its user agreement this week. Under its new third-party AI agent section, anything an agent you authorize does counts as authorized by you: its errors, its unintended actions, actions beyond the authority you gave. If the tool leaks your credentials or is hacked, what hackers then do is also treated as authorized. Citi says it has no duty to investigate, even when an agent conflicts with your prompts. Citi's own agents are exempt.
Who pays: the customer. Who is shielded: the bank. PNC changed its terms November 18; Citizens warned customers this week; JPMorgan and Bank of America, the two largest U.S. banks, moved in October. Agent makers are retreating — Anthropic's new finance tool is read-only. Whether the clauses survive federal liability limits is untested.
If your AI agent empties your account, should the loss be yours?