76% of AI agents sit dormant with live credentials: assign owners

Most AI agents in production are not doing anything. They still hold the keys. Arrakis Security's 2026 Autonomous Agents Threat Report says 76% of enabled AI agents have not run in 60 days, and their credentials and triggers still work. Help Net Security covered the report today. A dormant agent with live credentials is an unused employee badge that still opens every door. Nobody watches it, because nobody remembers it.

What actually happened: dormant AI agents with live access

Arrakis published the report on October 21. The data is anonymized production telemetry from thousands of agents in its customers' environments, collected over 12 weeks. That is vendor data, and it has not been checked by a third party. Read the numbers as a direction, not a census.

The key findings, from the Arrakis release and Help Net Security:

• 76% of agents had not run in 60 days. 44% had not run in 180 days.
• About 24% of dormant agents can be started by an unsigned webhook from outside the organization.
• 70% can reach sensitive systems such as cloud storage, AI services and marketing platforms. Almost none of those connections have verified data loss prevention.
• 20% can hand tasks to agents with different permissions.
• 82% carry at least one exposure. Help Net Security says 19% use API keys or tokens that never expire, stored in their configuration.

CEO Tal Baron put the cause plainly: when an agent stops being used, nobody removes anything, because nobody owns it.

Why it matters for your business

A small business does not have thousands of agents. It has eleven, and five were a pilot last spring. The Zapier flow with an OpenAI key. The n8n workflow on a webhook URL. The Make scenario that still has write access to Shopify. The support bot from the vendor you stopped paying attention to.

We treat agents like contractors. Every one gets:

• A named owner. One person. If that person leaves, the agent is reassigned or turned off that week.
• An expiry. Keys and OAuth grants rotate on a schedule. No token lives forever in a config file.
• A signed trigger. Every inbound webhook checks an HMAC signature or a shared secret. An open webhook URL is a public "run this agent" button.
• A kill date. An agent with no runs in 60 days is disabled automatically. If somebody misses it, they can turn it back on.

This costs an afternoon. Do the inventory first: list every automation platform, every API key in your vault, every webhook URL. Then delete what nobody can explain.

Sources: Help Net Security, Arrakis Security release via Hackread.