Over the past year or two almost every company has tried some kind of AI agent: a chatbot, an auto-reply for email, a flow that writes into the CRM. Many of them stopped running long ago. The question nobody asks: who took their keys away?
Arrakis Security’s 2026 Autonomous Agents Threat Report, covered by Help Net Security on 26 October, points straight at that blind spot. Over 12 weeks the company examined thousands of agents running in real production environments.
The numbers worth pausing on
- 76% of enabled agents had not run in 60 days; 44% had not run in six months.
- 70% are connected to at least one sensitive system (cloud storage, AI services, marketing platforms).
- 99.4% of agents with sensitive connections have no verified data loss prevention (DLP) coverage.
- 19% use API keys or tokens that never expire, embedded in their configuration.
- About a quarter of dormant agents can be started from outside through an unsigned webhook: anyone with the URL can trigger them.
- 82% had at least one recorded exposure; 46% had two or more.
Tal Baron, CEO of Arrakis Security, summed it up in one sentence: when an agent stops being used, nobody removes anything, because in most companies nobody owns it.
Why this matters for SMEs too
It is tempting to file this under “enterprise problems”. In a smaller company the pattern is even more common: someone sets up an automation in a no-code tool, gives it an admin-level API key to the web shop or the mailbox, the project fades, the colleague leaves, and the key stays. An attacker does not have to break anything; they only have to find the door that was left open.
The risk comes in three layers:
- Live access with no owner. A dormant agent reads and writes in your systems exactly as it did when you used it.
- Unverified triggers. If the webhook is not signed or IP-filtered, anyone can start the agent, with your permissions.
- Chained permissions. Agents hand tasks to each other and send data to external models. One misconfigured element opens the whole chain.
What to do this week
- Inventory. List every automation and AI agent: Make, Zapier, n8n, chatbots, voice agents, custom scripts. What is not on the list cannot be protected.
- An owner for every agent. One name who is accountable and decides when it gets switched off.
- Anything that has not run in 60 days: switch it off. Revoke its keys, delete its connections. If needed, it can be set up again in 10 minutes.
- Expiring keys, least privilege. No admin tokens: only the access the task needs, with an expiry date.
- Signed webhooks. Every trigger must verify where the request comes from.
- DLP on sensitive connections. At minimum, know what data leaves and where it goes.
How we build agents
At BoostYourBiz.ai, every AI Voice Agent and AI Chatbot rollout includes a permissions plan: each integration is created with the narrowest possible access, expiring keys and a documented owner, and stopped processes have their keys taken away. If agents are already running (or sleeping) in your company, the first step of our cybersecurity assessment is exactly this inventory.
Source: Help Net Security, “Forgotten AI agents may still access sensitive company data”, 26 October 2026 – Arrakis Security, 2026 Autonomous Agents Threat Report.