A report finds 76% of AI agents are idle but keep their access

On 22 October 2026, Arrakis released its 2026 Autonomous Agents Threat Report. The report finds that 76% of active AI agents did not operate in the last 60 days. Their credentials continue to work and they can continue to start tasks. The data is from thousands of agents in the systems of customers of Arrakis, in a period of 12 weeks. Arrakis makes tools for the protection of agents.

What it found

A credential gives access to a system. A webhook is an automatic signal from a web system that starts a task. Tal Baron, CEO of Arrakis, compares an employee and an agent. A company removes the access of an employee on the same day that the employee stops work. No person removes the access of an agent that stops. In most companies, no person is the owner of the agent.

The numbers

• 44% of agents did not operate in 180 days.
• 24% of idle agents can start from an unsigned external webhook.
• 70% of agents can get access to sensitive systems, such as cloud storage.
• Almost none of those connections have verified data loss prevention, which prevents data loss.
• 82% of agents have one or more exposures.

What is not known

The data is only from customers of Arrakis, and no other group examined it. The source gives no name of a customer. It reports no attack through an idle agent.

Insight: No person owns the agent, so no person ends its access

Arrakis tells that most companies have no owner for an idle agent. Salt Security found that 17.5% of organizations can stop an agent automatically. Nylas found that 66% of teams can fully examine agents that have an identity. This seems to be about owners, not tools. The first system to set an owner for each agent can become the control point.