AI agent permissions are already being rolled back inside large companies. A new survey from Agen.co by Frontegg finds that 78% of chief information security officers (CISOs) and chief security officers (CSOs) say their organisations have pulled permissions from an AI agent that was already running in production. The report, titled The State of Autonomous Agent Governance, was released on 27 October 2026.
Some respondents cut back access after a specific security concern, while others did it during a routine review.
What the survey says about AI agent permissions
The report is based on a survey of 300 CISOs and CSOs at US companies with 500 or more employees. It was fielded by the research firm Cint between 24 and 27 September 2026.
According to the report, 49% of respondents have found an agent running in production with access it should never have had. Only 37% scope an agent’s access to the task it was built for. That gap is why the company argues that the ability to narrow access after deployment has become a key governance requirement.
Agents are also busy. 84% of respondents see more than a thousand agent actions a week inside their production environments.
Confidence does not match incident rates
Three in four respondents (75%) say they are very confident they know how many AI agents are connected to their systems. Yet 52% of that confident group still had an agent security incident in the past 12 months, roughly the same as the overall sample.
Across all respondents, 54% reported a confirmed or suspected incident. Of those, 61% learned about it after the fact, from a log, an employee, a customer or an audit.
Among companies that reported an incident:
- 54% said the agent accessed data it should not have
- 43% said data was sent to an outside AI tool or model
- 14% said the agent kept running after it should have been stopped
Shared keys and borrowed identities
The single practice most linked to incidents in the survey was credential sharing. 53% of respondents run agents on shared or long-lived keys. Those companies reported a 60% security incident rate, against 46% for companies that do not. The report calls this the widest gap tied to any single practice.
Identity is another weak spot. 40% of respondents see agents running on an employee’s personal credentials, so the audit trail shows a person doing whatever the agent did. Only 18% verify an agent’s identity before it acts, and most of those check once, at registration or at the start of a session.
Ownership is also split. Security teams are held accountable for agent incidents at 48% of companies, but own the agents at just 23%. IT owns them at 41%.
What security leaders want next
The survey does not show security leaders trying to stop agents. Two-thirds want their organisations to keep adopting them, and most of that group plans to spend more on control. 86% plan to buy a dedicated agent governance tool within the next year, and more than half already have the budget.
Following agent incidents at leading AI labs this summer, 84% said they had tightened policy, added tooling or opened a review of how they manage agents.
Respondents also favour what the report calls bounded autonomy. 83% said they would trust an AI system to make agent access decisions in real time, as long as they set the limits. Only 23% would hand those decisions over entirely. Asked what would most reduce the effort of keeping agent policies current, the top answer was an AI that sets and updates policy on its own.
As a writer, I feel the most telling number here is not the 78%, but the 37%. If most agents are not scoped to their task on day one, pulling permissions later becomes routine clean-up rather than an exception. It is also worth remembering that Agen.co sells agent governance tools, so the survey comes from a company with an interest in the problem.
FAQ
What does the survey say about AI agent permissions?
78% of surveyed CISOs and CSOs say their organisations have pulled permissions from an AI agent already in production, and 49% have found an agent with access it should never have had.
Who was surveyed?
300 CISOs and CSOs at US companies with 500 or more employees, surveyed by Cint between 24 and 27 September 2026 for Agen.co by Frontegg.
Which practice was most linked to agent security incidents?
Running agents on shared or long-lived keys. Companies that do so reported a 60% incident rate, compared with 46% for those that do not.