As enterprise leaders focus on building out their agentic artificial intelligence programs, scrutiny over the security controls with AI agents across industries has been mounting.
ServiceNow published research last month covering the state of security for agentic AI and found that governance has not fully caught up with the speed at which organizations are deploying AI agents. In the past year, 90% of organizations reported an agentic AI security incident, but only 36% of them contained the incident within 24 hours.
“Confidence in AI tools is running ahead of the controls needed to manage them,” says Alexis Bliek, global head of risk, compliance and security GTM at ServiceNow. “In our study, 56% of leaders rate their agentic security as established, yet only 46% have security and governance controls in place. Look beyond that confidence, and the gaps widen: Only 20% have a single named, accountable owner for agentic AI governance, and only 19% can see more than 90% of their agents in real time.”
Bliek breaks down the report with lessons for financial services, as it included more than 1,000 senior security and technology decision-makers in the banking and insurance industries.
Click the banner below to learn more about how to improve automating organizational workflows.
Existing Gaps in AI Agent Accountability and Ownership
The gap between saying that agentic security has been established and having real security and governance controls in place is likely as prevalent in financial services as in other industries, Bliek notes, but it can have more serious consequences.
“Customers describe the core concern as being stuck in the middle and not knowing what they don’t know. Regulators also press on explainability, and when a customer asks why a mortgage application was declined, many teams do not yet have a good answer,” she says. “Financial institutions should not assume their existing safeguards will hold once agents start acting on their own.”
READ MORE: Why platform engineering may be the missing link in banking AI success.
Each AI agent should have a clear job description and boundaries, she notes, with an organization treating each one as a “registered asset.” That means it is registered, tested and approved before it goes live, closely monitored in production and easily stopped if it drifts.
A quarterly review would not be able to keep up with real-time development, so Bliek suggests a daily steering forum to keep leaders updated on deployments, risks and any anomalies.
“Agents are best treated as workers rather than software features. They need defined roles, access, monitoring and a way to be retired,” she says. “What is still missing from common understanding is that security is often added in after launch.”
Oversight must follow what AI agents actually do and have access to and not just their intended roles. Bliek also notes three areas that should have closer scrutiny:
- Data lineage. “Compliance and risk leaders increasingly need to show which data an agent used and where it came from.”
- Combined behavior. “Even when each model is sound, linking models can produce outcomes no one tested, so banks should test the connected system, not just its components.”
- Explainability. “Regulators want banks to show how automated decisions were reached and be able to explain them to customers. Every new tool, data set or system an agent can reach should trigger a fresh risk review rather than relying on a one-time sign-off.”
The Risk of Shadow Agents and the Lack of Visibility
For large financial intuitions, shadow AI agents can be a persistent problem, especially if one department knows exactly what an agent does but the IT team doesn’t have a record of it. Employees may also be relying on agents created by third parties.
34%
The percentage of organizations across industries that automatically register a new AI agent at deployment
Source: ServiceNow, “Deployed but Ungoverned: The State of Agentic Security,” October 2026
“Some banks are moving to shared operating models, where a partner runs operations on the bank’s platform and the customer adds its own AI layer. In that setup, three parties touch the same workflow, and visibility can break down at the seams,” Bliek adds.
DISCOVER: How to confront the new challenges of identity in the agentic era.
Inventory should then cover internal, vendor and partner agents. But, according to ServiceNow’s research, only 34% of organizations across industries automatically register a new agent at deployment, even though a majority say they have an accurate report of every agent in production. “Banks should make registration automatic at deployment, recording each agent’s owner, access and connected systems,” she says.
It’s clear that agentic AI security has room to grow across many sectors. In order for AI agents to act autonomously and be trustworthy, governance must improve.
“Before expanding an agent’s authority, leaders should ask teams to demonstrate how its current authority is controlled in practice,” Bliek says. “Before giving an agent more responsibility, prove you can manage the responsibility it already has.”
