Most dormant AI agents aren’t switched off. They sit idle with working logins, and that’s the main finding of a new security report. Arrakis Security says about 76% of enabled agents in its sample hadn’t run in 60 days, yet they kept their credentials and connections.
Think of it like a spare key you forgot you handed out. The agent does nothing for months, but the key still opens the door.
Here are the numbers worth knowing before we get into what to do.
- The report is Arrakis Security’s 2026 Autonomous Agents Threat Report, covered on October 26, 2026.
- It looked at thousands of agents in customer production environments over 12 weeks.
- About 76% of enabled agents hadn’t run in 60 days, and 44% hadn’t run in 180 days.
- 82% of agents had at least one recorded exposure.
Why idle AI agents are a bigger risk than they sound
An agent that never runs can’t do harm, right? Not quite. An idle agent still holds whatever access it was given on day one, such as cloud storage, AI services or marketing tools.
The trouble is ownership. Arrakis CEO Tal Baron put it plainly: “When an agent stops being used, nobody removes anything, because in most companies nobody owns it.”
That matches how teams work. Someone builds a helpful agent for a campaign or a test, the project ends, and nobody writes a ticket to shut it down.
The numbers behind the AI agent security problem
The table below pulls the main figures from the report, as relayed by Help Net Security.
| What Arrakis measured | Share of agents | Why it matters |
|---|---|---|
| Enabled agents that hadn’t run in 60 days | About 76% | Nobody is using them, but they still work |
| Agents idle for 180 days or more | 44% | Long-forgotten, rarely reviewed |
| Agents with at least one unsecured trigger | 22.9% | Anyone who finds the trigger could start the agent |
| Agents with credentials that never expire | 19% | Keys are embedded in the agent’s setup |
| Agents with at least one recorded exposure | 82% | Almost every agent had some weak spot |
These figures come from Arrakis’s own customer data, so treat them as a snapshot of that sample rather than the whole industry. Still, the pattern is easy to believe.
How an unused agent can still be started
The most worrying detail is the triggers. Arrakis found that about a quarter of dormant agents could be started through an unsigned webhook from outside the organization.
An unsigned webhook is a web address that accepts a command without checking who sent it. If someone finds that address, they may be able to wake the agent and use its access.
Credentials add to the problem. Around 19% of agents used API keys or tokens that never expire and were embedded in their configuration, so a leaked key stays useful for as long as the agent exists.
We’ve covered a related angle before: a hidden file can steer a coding assistant, as in the AGENTS.md attack. Forgotten agents are the quieter version of the same problem, because nobody is watching them at all.
What’s connected, and who is watching it
About 70% of agents were linked to at least one sensitive integration, such as cloud storage, AI or LLM services, or marketing platforms. Of those, 99.4% had no verified data loss prevention (DLP) coverage.
DLP tools flag sensitive data leaving a system. Without them, an agent could send data out and nobody would get an alert.
Some agents also had administrative privileges, or connections beyond their listed integrations. Others passed work to agents running under their own separate logins, which makes the access trail harder to follow.
A simple cleanup plan for dormant AI agents
The Help Net Security write-up doesn’t list fixes, so these steps are our own practical suggestions, not Arrakis’s recommendations. They work for a small team or a solo builder.
- List every agent you’ve created, including tests and demos, and note who owns each one.
- Switch off anything that hasn’t run in 60 days, then delete it if nobody objects.
- Rotate or revoke API keys on agents you’re keeping, and prefer keys that expire.
- Require a signature or secret on every webhook that can start an agent.
Why bother? Cleaning up takes an afternoon, while a stale key found by the wrong person can take weeks to untangle. For handling logins safely in the first place, see how 1Password for Agents lets AI log in without seeing your password.
Frequently Asked Questions
What is a dormant AI agent?
It’s an agent that is still enabled but hasn’t run for a long time. Arrakis counted agents as dormant when their last recorded run was 60 days ago or more.
Are dormant AI agents actually being hacked?
The report describes exposure, not confirmed break-ins. It shows that many idle agents could be reached or misused, which is the risk to fix.
Does this affect me if I’m not at a big company?
It can. Anyone who connected an AI agent to email, files or cloud tools and then forgot about it has the same gap, just on a smaller scale.
How do I find old AI agents I’ve forgotten?
Check each platform’s agent or automation list, then look at your connected apps and API key pages for entries you don’t recognize. Revoke anything you can’t explain.
Our take: treat every agent like a staff account. If it has no owner and no recent activity, turn it off today. You can always rebuild a small agent in minutes, but you can’t easily undo a leaked key.
Leave a Reply