More than half of organisations have experienced an AI agent-related security or privacy incident in the past year, according to new research highlighting significant gaps in enterprise security and governance.

The latest State of AI Agent Security report from AI agent management company Gravitee reveals that while businesses are rapidly expanding their use of autonomous AI systems, many lack the controls needed to monitor their behaviour, manage access and establish accountability when incidents occur.

Based on a survey of 900 senior technology leaders across the UK and US, the research found that 52% had experienced an AI agent-related security or privacy incident over the previous 12 months.

Despite this, only 17% of organisations assign individual identities to every AI agent, while fewer than one in ten have formally designated someone responsible for their actions.

Perhaps most concerningly, just 6% of respondents said their organisation could revoke a rogue AI agent’s access within an hour.

The findings come amid a sharp increase in enterprise AI deployment. The proportion of organisations operating more than 150 AI agents has more than doubled since April 2026, rising from 7% to 16%.

However, security capabilities have struggled to keep pace.

The proportion of technology leaders describing their organisations as very prepared to respond to an AI agent incident has fallen from 47% to 29% over the same period.

Meanwhile, the share of AI agents being monitored and secured has remained largely unchanged, increasing marginally from 56% to 57%.

Confidence in the regulatory environment has also declined, with just 26% of respondents now believing existing regulations adequately address AI agent risks, compared with 40% in April.

Accountability Concerns

The research also highlights challenges surrounding the oversight of AI systems that increasingly interact with other autonomous agents.

Almost all respondents (97%) reported operating AI agents capable of communicating with other agents, while 40% said their systems were connected to agents belonging to external organisations.

Despite these interconnected arrangements, only 27% had established clear accountability across the chain of interactions.

On average, organisations enforce policies governing permitted activities on just 51% of their AI agents, while only one in eight can comprehensively reconstruct an agent’s actions from 30 days earlier.

The consequences are already becoming apparent.

Some 58% of respondents reported having to revoke an AI agent’s access during a live security incident. However, just 25% subsequently held an identified individual accountable, while 17% made no changes following the incident.

Rory Blundell, CEO of Gravitee, characterised the findings as a failure of organisational oversight rather than technology.

“That is not a technology gap. It is an accountability gap, and it is widening as agent estates grow.”

Financial Services and Healthcare Among Most Affected

The research identified significant differences in AI agent security practices across industries, with some of the most concerning findings emerging from financial services and healthcare.

Despite demonstrating comparatively strong security controls, financial services organisations reported the highest incidence of AI agent-related security problems, with 66% experiencing an incident.

The sector also faces difficulties establishing responsibility, with 44% of respondents reporting uncertainty over accountability.

Healthcare organisations followed closely, with 63% experiencing an incident, including cases involving the exposure of patient data.

Recommended reading

• UK CFOs Upbeat on AI and Tech Investment for 2026
• UK Budget Unveils Major AI Investment Package
• Small Businesses Don’t Know Where to Start with AI Adoption
• Beyond The Hype: Turning AI Investment Into Business Value

Meanwhile, government and public sector organisations were found to operate some of the largest AI agent deployments, with a quarter running more than 150 agents.

However, just 8.8% of public sector respondents said they could fully reconstruct an AI agent’s actions, highlighting particularly weak auditability.

Telecommunications organisations also demonstrated shortcomings, with only 21% reporting a documented procedure for suspending or withdrawing an agent’s access.

In response to the findings, Gravitee has outlined an Agent Accountability Framework built around six principles: Identity, Sponsor, Role, Authority, Oversight and Recourse.

The framework is intended to help organisations establish which agent performed an action, the permissions under which it operated, who was responsible and how its activities could be stopped.

The report suggests that most organisations currently fall short of meeting at least one of these requirements.

The research was conducted by Opinion Matters in October 2026, surveying 600 senior technology leaders in the US and 300 in the UK across financial services, healthcare, government, telecommunications, manufacturing, and travel and transport.

The findings represent the third wave of Gravitee’s research, with comparisons drawn against its previous survey conducted in April 2026.