The spotlight is brightening on AI for cybersecurity as financial institutions navigate the technology’s growing risk-reward dynamic.  

Banks must view AI in cybersecurity as “a double-edged sword that radically compresses time for both the defender and the attacker,” Frances Zelazny, general manager of new market initiatives at identity verification platform Prove, told FinAi News. 

Prove works with 19 of the 20 largest banks, according to the company, including:

• Bank of America;
• Barclays;
• Citi;
• U.S. Bank; and
• Wells Fargo.  

“On the positive side, AI provides automated threat hunting by ingesting millions of log files, codebases and network packets to detect zero-day vulnerabilities and active intrusions in seconds rather than months,” she said.  “It also enables autopilot incident response.”

“If a breach occurs, a defensive AI agent can isolate affected servers, revoke compromised identity permissions and patch vulnerabilities autonomously, neutralizing attacks before human security analysts can even open an alert ticket.” 

— Frances Zelazny, GM of new market initiatives, Prove

Conversely, deploying AI for cybersecurity involves “severe operational vulnerabilities,” Zelazny said, highlighting the scale of the harm that can occur from a jailbreak when using powerful frontier models such as Anthropic’s Claude Fable 5.  

Attackers can manipulate enterprise AI agents via prompt injection, turning the tool against the bank to map out database vulnerabilities, she added.  

“There is also the threat of data poisoning and exfiltration,” she said. “To build effective defensive models, banks must feed them sensitive customer and financial data. This creates a massive target because if an attacker compromises the AI pipeline, they can distill or exfiltrate model data, effectively stealing proprietary banking intelligence or private customer identities.” 

Banks learning lessons 

More than 80% of banks cited cybersecurity as their top concern in 2025, according to a March report by financial services consulting firm Bank Director.  

Despite growing agentic AI capabilities for “always-on” cybersecurity, banks are recognizing when human-in-the-loop is needed, Peter Chapman, chief technology officer at Grasshopper Bank, told FinAi News.  

For example, the digital bank deployed a virtual AI assistant for cybersecurity, built through Google Gemini Enterprise, several months ago and realized that the agent was “as observant, if not more observant, than our human cybersecurity team,” Chapman said.  

“The thought there was we want to proactively leverage these frontier models to secure the border because bad guys are going to be proactively leveraging the frontier models to infiltrate the border,” he said.  

The agent continuously ran vulnerability scans, made recommendations and, initially, was allowed to execute remediations autonomously, Chapman said.  

But in one instance, the virtual assistant misclassified a production logging connection as a vulnerability and broke a logging service when remediating the issue, prompting the bank to tighten human oversight, he said. 

“It was relatively benign, but still, it’s not the type of thing you want to have happen,” he said. “So, we said to ourselves, ‘You know what, we skipped the human-in-the-loop approach for this.’ We learned from it, and we built in new processes, procedures and mechanisms to make sure that doesn’t happen again.” 

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas. This inaugural event will include speakers from Fifth Third and Capital One as well as a fireside chat with Piermont Bank founder and Chief Executive Wendy Cai-Lee.