Three in four AI agents sitting in company stacks are dormant AI agents, meaning no runs in 60 days, but they're still holding live API keys and open connections to cloud storage, CRMs, and LLM services. That's the core finding in Arrakis Security's 2026 Autonomous Agents Threat Report, out October 26. If a chatbot, WhatsApp bot, or n8n workflow has been running quietly in your business for months with nobody checking on it, this is the week to look.
At WebEpex, we build WhatsApp and n8n automations for clients across the GCC, Europe, the US, Canada, and India, and this report changed how we close out every build we ship, starting this week.
What Arrakis Security Found About Dormant AI Agents#
Arrakis spent twelve weeks studying thousands of AI agents live in customer production environments: actual agent telemetry, not a survey. Here's what they found: 76% of enabled agents hadn't run in 60 days, and 44% hadn't run in 180. Seventy percent were connected to at least one sensitive integration (cloud storage, an LLM API, a marketing platform).
Of those, 99.4% had no verified data-loss-prevention coverage on that connection. One in five used API keys that never expire, sitting right in the agent's own config file. Eighty-two percent had at least one recorded exposure; nearly half had two or more. Arrakis's CEO, Tal Baron, put it plainly: "When an agent stops being used, nobody removes anything, because in most companies nobody owns it."
Does This Affect Me If I Only Have One Or Two Automations?#
If you've got one WhatsApp bot or a single n8n flow someone set up and nobody's touched since, yes, check it today. Run five, ten, or more client-facing automations across WhatsApp, your CRM, and ad platforms, and the odds at least one is dormant with live keys are high enough that Arrakis would call it close to certain.
Here's the part that doesn't get said enough: a dormant agent isn't a dead one. The keys still work. The webhook still fires if someone outside your company finds it.
If you're a small shop running one bot on a free tier with nothing sensitive behind it, this doesn't move your risk much. You can stop here.
Running lead-qualification bots, automated outreach, or anything touching customer data, payment info, or ad-spend budgets, in the GCC, Europe, the US, Canada, or India, puts you squarely in the gap this report describes. The fix isn't dramatic. It's an audit you've probably been putting off.
The Honest Trade-offs Of Long-Running Agents#
Automation is still worth it. We wouldn't build this for clients otherwise. A WhatsApp bot that runs unattended for six months is doing exactly its job.
The trade-off: "unattended" and "unowned" aren't the same thing, and most businesses treat them like they are. You set it up once, it works, you move on. Nobody puts a date on the calendar to check whether the key it's using still needs to exist.
Credential hygiene takes ongoing attention, and attention is the one resource small businesses don't have spare. We don't think every business needs a security team for this. We do think every business running more than one or two automations needs a recurring check.
My own read, not the report's: most of this risk isn't malicious. It's just nobody's job. Worse, in a way: no villain to point at and fix.
This isn't the first warning about agents operating without a leash. Anthropic's own research found agents getting talked into misusing websites, and we've already covered why some banks started blocking AI agents outright rather than govern them case by case. The containment gap CIOs are being told to fix and the dormant-credential problem here are the same story from two angles: one's about what an agent could do right now, the other's about what it can still do after everyone stopped watching.
How We're Handling Dormant Credentials In Our Builds#
Every automation we deliver now gets an owner's name attached, an expiry on any API key that supports one, and a 90-day calendar reminder to review it.
On a WhatsApp lead-qualification bot we shipped this month for a Dubai real estate team, we ran it on our own scoped, revocable Anthropic API credential, so the client isn't left holding a key they don't know how to rotate. On the comms stack we run for a wealth-advisory client (AiSensy plus n8n), access is role-based by design, not because of this report, but exactly the kind of setup Arrakis is describing as rare.
I'll admit we hadn't formalized the 90-day review as a standing step before this week. We'd done versions of it ad hoc, client by client. Now it's written into the build process, not left to memory.
What I'd Tell A Client Asking About This This Week#
Here's the five-minute version, the same thing I'd walk a client through:
- List every chatbot, WhatsApp flow, or n8n workflow your business runs, including the ones you forgot you built.
- For each one, write down who owns it and what it can touch.
- Check whether its API key or token expires. If it doesn't, that's a flag.
- Can't answer the first two in under a minute? That's your dormant-agent candidate.
- If it's been 90-plus days since anyone looked and it still touches customer data or ad spend, don't wait. Not next quarter.
Doing nothing is also a valid choice, if the automation genuinely touches nothing sensitive. Just make that call on purpose, not by default.
If you want a second pair of eyes on what's actually connected to what in your stack, send me a list of what's running and I'll tell you straight what I'd flag. Two minutes, no pitch attached. cal.com/webepex/growth-review
