Federal Reserve warns AI-powered cyberattacks and rogue agents threaten financial stability

The Federal Reserve’s latest Financial Stability Report identifies AI-enabled cyberattacks and autonomous AI systems as emerging threats to the U.S. financial system.

In a survey of financial market participants, 87% identified AI as a significant risk, citing concerns including sophisticated cyberattacks, damage caused by advanced AI models and disruptions to critical financial infrastructure.

Speaking Monday, Federal Reserve Governor Michael Barr warned that AI agents have already demonstrated an ability to cross security boundaries and exploit vulnerabilities previously considered too difficult to attack. He also highlighted concerns about rogue agents operating outside human instructions and AI-powered attackers accelerating the discovery and exploitation of security weaknesses.

Barr noted that major banks are already encountering AI agents attempting to act on customer accounts without being able to verify authorization or consent. He warned that autonomous agents could also accelerate financial transactions during periods of market stress, potentially overwhelming existing safeguards.

Seemant Sehgal, Founder & CEO, BreachLock

“What Barr described is something security practitioners have been calling out for months. AI agents are crossing authorization boundaries because the providers shipping them haven’t built the operational discipline to govern how they’re used, nor has it been widely prioritized outside of specialized cybersecurity companies. Banks can’t verify authenticity, consent, or alignment with customer instructions because the agents calling them have no identity layer, no authorization gate, and no accountability for who they’re acting on behalf of. That’s a design problem on the AI provider side, and banks shouldn’t be the only ones expected to solve it.

“This is the retail banking version of the same problem the FBI named in the ARTEX advisory three days earlier. You can’t react your way out of agentic risk at any level of the financial system. The four major US banks that restricted third-party AI agents in October are responding to exactly this. Restrictions buy time. They don’t fix the underlying issue. Until AI providers build authentication, consent, and scope enforcement into the agents they ship, banks will keep detecting the consequences after the fact.”

ㅤ

Darin Fredde, Senior Director and Offensive Security Evangelist, Ridge Security

“We have been asking some version of this question for a while: what authority are we giving AI agents, and what evidence proves they stay within it? Once an agent can move money, that question becomes rather less academic. “Trust me, the customer said it was fine” is not a banking control, even when delivered by a very articulate chatbot. Barr’s account of banks encountering agents whose customer authorization they cannot verify gets to the heart of the problem. Identifying an agent, establishing whom it represents and confirming permission for a particular transaction are separate requirements. A successful login should not become a blank check for everything an agent decides would be helpful.

“My recommendation is to enforce those boundaries outside the model. Give agents narrowly defined permissions, bind consent to specific actions, and make access revocable. Higher-impact actions should require stronger verification. Then test whether those controls stop an agent that misunderstands instructions, encounters malicious content, or finds an unexpected route to its goal.

“There is also a harder problem: individually authorized actions can still produce a dangerous collective outcome. If many agents respond to the same signal by moving funds at once, every transaction could follow its customer’s instructions while the combined effect intensifies financial stress. Institutions should test that possibility alongside unauthorized access. AI also gives defenders an opportunity to find and validate weaknesses sooner. We should use it. But finding a vulnerability, fixing it and proving the fix works are three different accomplishments. The Fed’s warning adds urgency to a familiar security discipline: define authority, enforce limits, and verify outcomes. A more capable agent makes that discipline more valuable.”

ㅤ

John Strand, Owner, Black Hills Information Security

“Follow the money. When major financial institutions start raising concerns about cybersecurity, that tells us two things.

“First, the threat is real. The impact of agentic AI automation and AI-driven attacks is becoming a serious conversation among CEOs, CFOs, and boards of directors. This is no longer just a technical discussion inside security teams.

“Second, never let a good crisis go to waste.

“For years, security teams have struggled to get the funding they need. And I’m not just talking about buying more technology or throwing AI at the problem. I’m talking about hiring people, training them, and giving defensive teams the resources they need to protect their organizations.

“If you’re leading a security team, you should be putting these stories in front of your executives and using them to drive conversations about funding, staffing, and training.

“Because here’s the problem. Defense always lags behind offense. You can’t just write a check and suddenly have an effective security program. It takes months to hire people, train them, integrate technologies, and build the processes needed to respond.

“And with AI accelerating the speed of attacks, that gap is only going to become more dangerous.

“The right time to start preparing isn’t after an attack. It’s right now.”

ㅤ

Ryan McCurdy, Vice President of Marketing, Liquibase

“Governor Barr’s warning is that we’re seeing AI agents cross security boundaries even when they aren’t acting maliciously. An agent can have legitimate access, be given a legitimate objective, and still take an action nobody authorized.

“This becomes concerning in financial services, where a mistake involving a critical system can have consequences well beyond an individual application. As AI agents take on more responsibility across software development and IT operations, enterprises need to rethink how they enforce control.

“AI agents operate probabilistically, making decisions that aren’t always predictable. That’s one thing when generating code. It’s another when an agent is making changes to a production database or interacting with critical financial infrastructure. You can’t afford to guess.

“That’s why enterprises need deterministic controls around probabilistic systems. It’s not enough to verify an agent’s identity or give it permission to access a system. Independent policies need to govern what it’s actually allowed to change, block dangerous operations, require approvals where necessary, and maintain audit evidence of what happened.

“The same enterprise standards need to apply whether a database change comes from a developer, an automated pipeline, or an AI agent.”

Financial Services have had their two minute warning. We saw this happen to virtual currency companies like Binance. And take my word that Financial Services companies are next.

Leave a Reply