<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0"><channel><title>Cooper Beckwith Research</title><link>https://cooperbeckwith.com/research/</link><description>Papers on AI risk, agents, capital, and liquidity in financial institutions.</description>
<item><title>Amex Bank Just Agreed to Pay $350MM for AML Risks It Didn't Know It Had. What Happens When the AI Agents Get Involved?</title><link>https://cooperbeckwith.com/research/amex-350mm-aml-and-ai-agents</link><guid>https://cooperbeckwith.com/research/amex-350mm-aml-and-ai-agents</guid><pubDate>Wed, 14 Oct 2026 06:30:00 -0400</pubDate><description>On October 8, 2026 the OCC found systemic breakdowns in American Express National Bank's suspicious activity monitoring. Its risk assessment watched deposit products while about $13 billion of suspected trade-based money laundering moved through card charges and repayments over eleven years. A BSA and AML program watches what its risk assessment tells it to watch, and AI agents are the next business most risk assessments have not seen.</description></item>
<item><title>Forty-Five Minutes, $460MM, and 97 Unread Emails</title><link>https://cooperbeckwith.com/research/forty-five-minutes-460mm-97-unread-emails</link><guid>https://cooperbeckwith.com/research/forty-five-minutes-460mm-97-unread-emails</guid><pubDate>Mon, 12 Oct 2026 06:30:00 -0400</pubDate><description>On August 1, 2012, Knight Capital lost more than $460 million in about 45 minutes to trading code it had retired years earlier and never removed, while 97 automated warning emails reached staff who did not act on them. The firm had no procedure to stop its own machine. The agent you deploy is the agent you own, and contract law, a 2024 tribunal, and a bipartisan Senate bill announced October 1, 2026 all say so.</description></item>
<item><title>The Credentials Were Valid. The Instructions Were Not.</title><link>https://cooperbeckwith.com/research/the-credentials-were-valid</link><guid>https://cooperbeckwith.com/research/the-credentials-were-valid</guid><pubDate>Fri, 09 Oct 2026 06:30:00 -0400</pubDate><description>In February 2016, $951 million of fraudulent payment instructions sent from Bangladesh Bank's own SWIFT terminal passed authentication, and $81 million was stolen. Authentication proves who holds the key. Authorization proves the holder had authority for this instruction, at this amount, to this payee, today. A customer's AI agent will present perfectly authenticated instructions, and the control that tells a legitimate one from a stolen or mistaken one is a record of what the customer authorized: the mandate.</description></item>
<item><title>The New Casas de Cambio</title><link>https://cooperbeckwith.com/research/the-new-casas-de-cambio</link><guid>https://cooperbeckwith.com/research/the-new-casas-de-cambio</guid><pubDate>Wed, 07 Oct 2026 06:30:00 -0400</pubDate><description>Between you and the person whose money moves there is now a platform that runs the agent: a frontier lab's assistant, a big tech wallet, a fintech, a data aggregator, or an agent sold through your core provider's marketplace. That platform is the operator. It identifies its own principals, aggregates thousands of them, and presents your institution with volume at machine speed. Wachovia paid $160 million in 2010 for treating $420 billion of casa de cambio traffic as if it came from customers it knew.</description></item>
<item><title>Who Is Behind the Agent at Your Window?</title><link>https://cooperbeckwith.com/research/who-is-behind-the-agent-at-your-window</link><guid>https://cooperbeckwith.com/research/who-is-behind-the-agent-at-your-window</guid><pubDate>Tue, 06 Oct 2026 06:30:00 -0400</pubDate><description>An AI agent can open a session at your institution with your customer's credentials, read the account, and move money, and to your systems it looks like the customer. In 2012 HSBC paid $1.92 billion for knowing its counterparty and not the principals behind the counterparty's traffic. The agent at your window opens the same gap.</description></item>
<item><title>AI Risk in Depositories: The Need for Elicitability</title><link>https://cooperbeckwith.com/research/ai-risk-in-depositories</link><guid>https://cooperbeckwith.com/research/ai-risk-in-depositories</guid><pubDate>Thu, 17 Sep 2026 06:30:00 -0400</pubDate><description>Artificial intelligence is real, compounding, and about to accelerate as models begin improving themselves. That is not a reason to run it into your books and records, and it is not a reason to stand in front of it. It is a reason to be deliberate. From 2006 to 2008, opaque models stacked on leverage produced numbers no one understood and losses no one could locate until it was too late, and community institutions carried real damage.</description></item>
</channel></rss>
