Fed Governor Barr Says AI Agents Could Accelerate Bank Runs That Unfold in Minutes
A.I. / news
Fed Governor Barr Says AI Agents Could Accelerate Bank Runs That Unfold in Minutes
In a Nov. 16 speech in Milan, Barr also put AI firms' off-balance-sheet liabilities above $3 trillion and proposed no new rules.

Federal Reserve Governor Michael S. Barr said on Nov. 16 that AI agents could move deposits fast enough to accelerate bank runs "that unfold in minutes", faster than banks' liquidity management or regulators can react.
He made the remark in a speech at the Association for Computing Machinery's International Conference on Artificial Intelligence in Finance in Milan. The Fed published the text the same day under the title "Some Potential Financial Stability Implications of Artificial Intelligence". Barr said the views were his own and not necessarily those of the Board or the Federal Open Market Committee.
What Barr said about AI agents
In the speech, Barr said AI agents "could instantaneously detect yield differentials or stability signals and autonomously move funds at high speed". Where customers once needed to execute transfers themselves, he said, that could overwhelm "traditional liquidity management and regulatory intervention timelines".
He also said major banks now report finding AI agents attempting actions in retail customer accounts, without the bank being able to verify authenticity, consent and alignment with customer instructions. He did not name the banks or give a count. Agent-run transactions are already a product category, as OpenAI's Odessia acquisition shows.
Barr proposed no new rules. He called for strict authentication, explicit consent checks for specific actions, technical limits on agent capabilities and both rules-based and values-based constraints. Gizmodo cited an October Axios report on banks finding agents trying to reach customer accounts without authorization.
The financing figures in the same speech
Most of the speech concerned how AI buildouts are paid for, a theme in coverage such as the Anthropic IPO range report. Barr said on-balance-sheet debt of the seven largest AI firms is expected to exceed $800 billion before the end of 2026.
He said the Fed's Financial Stability Report estimates that off-balance-sheet liabilities, such as long-term contracts, lease commitments and residual value guarantees, now exceed $3 trillion in potential maximum nominal value. Firms generally report these before discounting, so the figure is not directly comparable with the debt number.
| Item in Barr's speech | Figure |
|---|---|
| On-balance-sheet debt, seven largest AI firms | Above $800 billion by year end |
| Off-balance-sheet liabilities, potential maximum | Above $3 trillion |
| Equity investments in LLM labs and cloud start-ups | Above $100 billion cumulative |
The $100 billion figure covers hyperscalers and GPU makers investing in LLM labs and cloud start-ups. Barr called that a form of circular financing, because the smaller firms use the money to rent or buy compute from the investor.

Other risks he listed
On trading, Barr said similar models trained on similar data could produce herding, with strategies that synchronise their reactions to the same news. He said that, to his understanding, generative AI is not yet used for large-scale autonomous trading.
On cyber risk, he said AI models "have treated security boundaries as obstacles to overcome rather than rules to be followed". Cyber offense, he said, "can experiment and fail repeatedly until something works", while defenders must be right all the time.
Gizmodo, citing a New York Times report of Nov. 16, said Moody's counts $85.2 billion in AI-related debt held by Athene and Corebridge Financial as of June, and that Barclays estimates life insurers own about 15% of AI data centre debt. Those figures come from the New York Times as relayed by Gizmodo, not from Barr's speech.
The speech gives no date for any supervisory action. It is Barr's fourth speech devoted to AI in the past year, he said.
Sources
More in A.I.
- 01Runta's 27,280-Run Harness Eval Finds the Model Explains 92.8% of Agent CostFrontierHarness v1.1 ran 22 coding-agent harnesses against five open models, and Codex CLI went from first place to sixteenth when the model changed.
- 02Qwen3.8-27B Has 6.9 Million Downloads on Hugging Face. Artificial Analysis Rates It Slow and VerboseAlibaba's Apache 2.0 dense model beats its size class on an independent index of 20, and takes 38.8 minutes a task against a 12.6-minute median.
- 03Claude Code Team Publishes Its Verification Rules: 39 Checks, a Three-Round Limit and Observe ModeA Nov. 16 post says about 90% of the team's code is written and reviewed by Claude, and lays out the gates that keep it shippable.
- 04Inkling 2 Small Weights Land on Hugging Face Under Apache 2.0, and the Card Asks for 500GB of VRAMThinking Machines opened the weights a week after launching on Tinker. The safety section still says release is pending.