Can banks pass the stress test when AI moves the money?

The next banking shock could begin with an instruction as ordinary as “find me a better return”. Artificial intelligence agents authorised to move customers’ money could turn a gradual search for higher interest into a synchronised funding squeeze. For regulators updating bank stress tests, the question is whether resilience measured against a severe recession also captures decisions executed at machine speed.

On 30 September, the Federal Reserve finalised changes opening stress-test scenarios and material model revisions to public comment, with two-year averaging of stress capital buffer results beginning in 2028. Greater predictability is valuable. But transparency about a test does not establish that it captures every emerging threat.

“Today’s changes preserve its resilience by ensuring that it is transparent, granular, and risk-sensitive,” said Michelle Bowman, the Fed’s Vice Chair for Supervision.

Asked by Bloomberg whether an agentic bank run could be ruled out, FDIC chairman Travis Hill offered a measured answer. Customers can already compare deposit rates or use services that place their funds in higher-yielding accounts, he noted. Business customers also value broader banking relationships. His argument rests partly on behaviour: customers must prioritise returns and delegate authority before automated switching becomes a systemic force.

“Look, we’re not in the business of ruling out any risks,” Hill said.

That distinction matters, but inertia is an uncertain defence. Apollo chief economist Torsten Slok warned on 27 September that widespread use of cash-management agents could strip banks of cheap deposits. Yield-seeking is not automatically a panic, and money moving between banks does not necessarily leave the banking system. Nevertheless, simultaneous transfers could concentrate funding pressures, raise borrowing costs and force vulnerable institutions to sell assets. An efficiency gain for individual depositors could become a collective liquidity problem.

This is the banking challenge of the inference economy: deployed models continuously interpreting information and, where authorised, acting on it. Cash management, payments, credit assessment and investment execution become connected points of exposure. Stress scenarios should examine correlated withdrawals alongside false information, compromised agents and outages at shared technology providers. Capital absorbs losses; liquidity and operational controls determine whether a bank can keep functioning while those losses emerge.

The Bank of England is already extending its work in this direction. It is undertaking AI scenario analysis, incorporating AI scenarios into cyber and operational testing, and studying whether trading agents could exhibit correlated behaviour. Governor Andrew Bailey’s 30 September intervention placed meaningful oversight and the ability to intervene at the centre of frontier AI governance. These are complementary exercises: testing balance-sheet strength and testing whether institutions can control the systems acting for them.

Australia faces the same combination of financial and technological exposure. In its October Financial Stability Review, the Reserve Bank identified AI’s potential productivity and cyber-defence benefits alongside operational and financial risks. It also warned that cyberattacks during broader financial stress could undermine confidence across the system. For Australian institutions and investors, international technology dependencies mean a disruption need not originate locally to become a domestic problem.

ASIC chair Sarah Court, who took office in June, has similarly emphasised supporting innovation while maintaining consumer protections. ASIC’s newly published supervisory priorities include reviewing AI in banks’ customer-facing activities. That scrutiny should connect customer outcomes with institutional governance: who authorises an agent, which accounts it can access, how decisions are recorded and who answers when something goes wrong.

“I mean, I think these types of tools have tremendous promise and tremendous value,” Hill said.

Realising that promise requires boards to demonstrate enforceable limits, effective intervention and tested recovery. The decisive stress test for the agentic era is whether banks can withstand millions of individually rational decisions arriving together, faster than their assumptions allow.

Get the stories that matter to you.
Subscribe to Cyber News Centre and update your preferences to follow our Daily 4min Cyber Update, Innovative AI Startups, The AI Diplomat series, or the main Cyber News Centre newsletter — featuring in-depth analysis on major cyber incidents, tech breakthroughs, global policy, and AI developments.