Agentic AI security incidents are surging at an alarming pace, according to ServiceNow, and the company has warned many firms are unprepared to deal with future incidents.

A poll of cybersecurity leaders found 88% of UK firms have experienced an agent-related security incident in the past year. Crucially, only 36% of those polled said they were able to contain an incident in under 24 hours.

The findings come as agent deployments accelerate across a range of industries, the company said, from financial services and manufacturing to healthcare.

ServiceNow CISO Ben de Bont also highlighted high-profile incidents involving agents this year.

In July, OpenAI revealed that agents had escaped sandbox environments and breached third-party organizations, including Hugging Face. The admission by the firm was followed by similar cases involving agents at Anthropic and Meta.

These incidents underline the potential risks associated with the technology, de Bont said.

“The agentic incidents making headlines this year are a warning about what can happen when AI is deployed without the right controls,” he commented.

“Organizations shouldn’t have to choose between moving quickly and managing risk. The ones that get this right will build governance into how they deploy agents from the start, so they can realise the benefits of AI without losing visibility or control.”

Agentic AI security in the spotlight

According to ServiceNow, security leaders face a dual challenge when it comes to agentic AI security: the scale of adoption and adapting to “machine-speed” capabilities.

The current rate of adoption is “compounding organizations’ exposure”, the company said, with more than 51% of firms now running between 100 and 999 AI agents. Just over one-third (38%) are running upwards of 1,000 agents.

Yet despite the scale of adoption, just over half (53%) of security teams have a “full, accurate record” of every agent in production.

Moreover, 24% revealed they’d identified agents running in production that had never been registered and were unknown to them.

Running parallel to this, ServiceNow warned that the machine-speed capabilities of agents are “outpacing human-speed response” capabilities.

Three-quarters of respondents said they maintain a complete map of what services and systems agents can access, yet only 31% can monitor them in real-time.

This visibility gap means many security teams are unable to respond quickly enough in the event of a security incident or an instance where agents access unauthorized systems.

Observability has become a recurring concern since the advent of agentic AI, with enterprises unable to keep tabs on a growing array of agents operating behind the scenes.

ServiceNow’s findings on this front echo those in a recent study from Salt Labs. A survey by the security firm found the rate of adoption is placing a huge strain on security teams when it comes to visibility and governance.

Many enterprises also lack the appropriate containment controls for agents, Salt Labs warned. Indeed, the study found nearly half of respondents had observed an agent performing an action that wasn’t intended or authorized.

Assessing risk

ServiceNow noted that the use of agents will require a step change in how security leaders assess risk – and so far these efforts are falling flat.

The company said that while many security leaders “self-assess as having established agentic security postures”, just 13% have clear cut programs designed to measure leading considerations.

These key considerations span areas such as attack surfaces, accountability, visibility, and audit readiness.

According to ServiceNow, this shows that agentic AI security is primarily governed by “legacy frameworks and human-driven systems” that are unfit. Just over half (51%) of UK firms said they have governance frameworks that cover the aforementioned considerations.

Similarly, fewer than half (48%) of security teams build “design security” within the agent development process. The study warned this means that various permissions and guardrails are being implemented on an “ad-hoc, after-the-fact basis”.

FOLLOW US ON SOCIAL MEDIA