Financial institutions must ensure that agentic AI models have a “kill switch” at multiple levels to prepare for regulatory changes and the inherent risks of autonomous AI, experts say.  

More than one-third of U.S. banks lack essential kill switches — safety mechanisms that pause or disable AI applications — according to a May 28 report by professional services firm Wolters Kluwer, which surveyed 230 industry professionals.  

These tools are especially crucial for AI agents because the technology acts autonomously, Vishal Dalal, chief product officer for financial solutions at fintech Fiserv, told FinAi News.  

“If an agent starts going in a loop, starts making decisions that it shouldn’t, there has to be a kill switch,” he said.  

In a June 30 speech at the European Central Bank forum, Bank of England Deputy Governor Sarah Breeden called for such mechanisms in case autonomous trading bots cause a “market meltdown” through unintended “herding behavior.” 

Task, process, model layers 

Most agentic models should have three basic kill-switch levels, each with varying degrees of use depending on the level of autonomy, Joshua Summers, chief executive of agentic lending platform EnFi, told FinAi News.  

Those levels are:

1. Task level: Stopping simple processes for basic reasons; 
2. Process level: Stopping an entire workflow to understand an agent’s reasoning, update audit logs or improve overall model performance; and
3. Model level: Shutting down an AI system altogether because of high-risk situations, whether it’s bad actors, security breaches or agents going astray.  

Anthropic‘s Claude Fable 5 is one type of model that necessitates the strongest safety protocols because of its immense power, Summers said. 

The U.S. government temporarily disabled Fable 5 three days after its June 9 launch when researchers identified a “jailbreak” method to bypass the model’s cybersecurity safeguards, according to Anthropic.  

“I think a lot of those exist; we just don’t see them,” Summers said. “It is critical to think where this all goes as AI is given a seat at the table.” 

For now, EnFi is not concerned about “rogue agents” with its model because “we’re not building a decisioning system that runs on its own,” Summers said. 

“That’s not to say we don’t eventually go past that and get to maybe more automated judgment calls and decisioning, but that’s not where we are in credit today.” 

Preparing for regs 

The potential for legislation that prohibits certain agentic applications is another reason for kill-switch protocols, Ash Govindia, senior vice president of growth at AI software platform FintechOS, told FinAi News.  

For example, lawmakers could enact a rule that says, “stop letting AI chatbots talk to the customer through a mortgage application,” Govindia said. 

“I would consider that a kill switch and you should be able to turn off that scenario where the agent prompts [borrowers] for three more months of pay stubs and stuff like that,” he said.  

There also could be more instances like the Fable 5 shutdown, with agencies banning or restricting access to certain models, Govindia said. 

Ultimately, kill switches must be viewed as a “base functionality” for any AI model, he said.  

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas.