Enterprise AI Agent Fleets Expand as Security Governance Lags

By Amara Osei

The share of surveyed organizations operating more than 150 AI agents rose from 7% in April 2026 to 16% in October 2026, according to Gravitee. The company released the third wave of its State of AI Agent Security survey on Monday.

The survey covered 900 senior technology leaders: 600 in the United States and 300 in the United Kingdom. Opinion Matters conducted the October 2026 fieldwork, and Gravitee commissioned the study, which compares results with its April 2026 wave.

What changed between April and October?

Only 17% of respondents said they give each agent its own identity. Fewer than one in 10 have a named person formally accountable for what their agents do.

“Agents are making decisions and acting on their own inside enterprises, including banks, hospitals and government departments,” said Rory Blundell, CEO of Gravitee.

“When one goes wrong, most organizations cannot say which agent it was, what it was allowed to do, or who is responsible. That is not a technology gap. It is an accountability gap, and it is widening as agent estates grow.”

How prepared are enterprises to contain agent incidents?

Gravitee’s survey recorded falling confidence as deployment expanded. The share of respondents who said they were “very prepared” to respond to an agent incident fell from 47% to 29%, while belief that current regulation covers agent risks dropped from 40% to 26%. The proportion of agents monitored and secured barely moved, from 56% to 57%.

More than half of respondents, 52%, reported an agent-related security or privacy incident in the past year. Separately, 58% said they had revoked an agent’s access during a live incident. After doing so, only 25% held a named person responsible, and 17% changed nothing.

Where do agent controls break down?

Gravitee’s survey measured three acute shortfalls in policy enforcement, historical reconstruction and emergency intervention. The results were:

• Policy enforcement reaches only 51% of agents on average.
• Only one in eight organizations can fully reconstruct what an agent did 30 days ago.
• Only 6% can revoke a rogue agent’s access within an hour.

Nearly all respondents, 97%, said they run agents that call other agents, and 40% connect to other organizations’ agents. Yet only 27% define who is accountable across that chain.

“Every organization we speak to wants to scale agents, and none of them want to slow down,” said Linus Håkansson, CPTO of Gravitee.

“The answer is to manage agents the way you would manage any employee with access to sensitive systems. Give each one an identity, enforce what it can do, keep a record, and maintain the ability to revoke access within minutes when something goes wrong.”

Which sectors report the sharpest gaps?

The survey’s sector breakdown showed:

• Government and public sector: These organizations had the largest fleets, with 25% running more than 150 agents. Only 8.8% could fully reconstruct an agent action.
• Financial services: The sector reported the strongest controls but also the most incidents, at 66%. Ownership remained least clear, with 44% saying accountability was unclear.
• Healthcare: 63% reported an incident, including exposed patient data.
• Telecommunications: Only 21% had a documented process to pause or revoke an agent.

What standard has Gravitee proposed?

Gravitee’s Agent Accountability Framework sets six principles that the company says every agent in production should meet: Identity, Sponsor, Role, Authority, Oversight and Recourse. The research says most organizations fall short on at least one.

Gravitee, a vendor of AI agent management software, sells the Gamma platform. The product spans Agent Management, API Management, Event Management, Authorization Management, and Identity & Access Management.

Blundell will discuss the findings at a Gravitee event at the National Gallery, London, on November 18. The full report is available at gravitee.io/state-of-ai-agent-security. For enterprises expanding agent fleets, identity, authority, recordkeeping and minute-level revocation now form the practical test identified by Gravitee.