🎧 Prefer to listen?

I’ve been running AI agents to automate parts of my workflow — scheduling, content drafts, data pulls — and until last week, I never thought twice about what permissions those agents actually have. Then I read a study that made me audit every single one.

VentureBeat surveyed 107 enterprises and found that 54% have already had a confirmed AI agent security incident or a near-miss. Not “concerned about the possibility” — already happened. And this is in companies with dedicated security teams and budgets. If it’s happening to them, it’s definitely happening to us.

I wrote about how AI agents are becoming your next digital employees — but I didn’t spend enough time on what happens when those employees go rogue. Let me fix that.

What the research actually found

The numbers are stark:

  • 54% of enterprises have experienced a confirmed agent incident (18%) or a near-miss caught before damage (36%)
  • Only 32% give every agent its own scoped identity — the rest share credentials across agents
  • Only 30% isolate their highest-risk agents in sandboxes
  • Companies with credential sharing hit at 63.5% incident rates vs. 40.9% with fully scoped agents

The core problem is identity. Most organizations are running multiple AI agents on shared API keys or human credentials. When one agent gets compromised or over-permissioned, the blast radius covers everything that key touches.

I explained how tool calling actually works in a previous post — but knowing how agents execute actions is only half the picture. The other half is understanding what they can access while doing it.

Why this matters even if you’re a team of one

You might think this is enterprise-only concern. It’s not. Here’s why:

You’re probably running agents on shared keys right now. If you’ve connected ChatGPT to your Google Drive, or given Claude access to your email, or set up a Make.com scenario that touches your calendar, your CRM, and your invoicing — those are all running on your personal OAuth tokens or API keys. One misconfigured agent has access to everything.

Solo builders skip the boring stuff. Enterprise security teams at least have checklists. Most of us set up an automation, see that it works, and move on. We don’t audit permissions, we don’t scope access, and we definitely don’t set up sandboxing. The groupthink problem I’ve written about before applies here too — we assume because everyone’s doing it this way, it must be safe.

The damage is proportional to your digital footprint. For a solo builder, a compromised agent that has access to your email, your social accounts, your payment tools, and your client data isn’t just a security incident — it’s a business-ending event.

The five things I changed after reading this

These aren’t theoretical. I did all of them this week.

1. Audit every agent’s permissions

Go through every AI tool and automation you’ve connected. For each one, ask: what can this agent actually access? If you gave ChatGPT access to your Google account, does it have read-only access to Drive, or full read-write-delete? Most people don’t know.

I found three automations running with broader permissions than they needed. Took 20 minutes to fix.

2. Stop sharing API keys across tools

If you’re using the same OpenAI API key for your chatbot, your content automation, and your data analysis tool — stop. Create separate keys for each use case. If one gets compromised, you can revoke it without killing everything.

This is the solo-builder version of “scoped identity.” It costs nothing and takes five minutes per key.

3. Use read-only where possible

Not every agent needs write access. If you’ve connected an AI to your email to summarize it, it doesn’t need send permissions. If it’s reading your calendar to plan your day, it doesn’t need to create events.

I went through my OAuth connections and downgraded anything that only needed to read. The principle is similar to what I covered with tool calling — agents should have the minimum permissions to do their job, not maximum permissions because it was easier to set up.

4. Set up alerts for unusual activity

Most AI platforms let you set usage alerts. If an agent that normally makes 50 API calls a day suddenly makes 5,000, you want to know about it. I set up spending alerts on every API key I use.

For Make.com and Zapier, I added email notifications for scenario failures. A failing automation that keeps retrying can rack up costs or corrupt data before you notice.

5. Isolate your highest-risk automations

The VentureBeat study found only 30% of enterprises isolate their highest-risk agents. For solo builders, “isolation” might mean: run your payment-related automations on a separate API key from your content automations. Use a dedicated email address for agent connections rather than your primary inbox.

If you’re running agents that touch money, client data, or publishing — treat those differently from agents that just summarize articles or draft social posts.

The tooling gap nobody’s talking about

One finding from the study stood out: enterprises are overwhelmingly using provider-native security (OpenAI’s guardrails at 51%, Google and Microsoft’s cloud controls) rather than purpose-built agent security tools. And satisfaction is high — 4.2 out of 5.

But here’s the catch: the same companies report they plan to change their tooling within the year. They’re satisfied with controls they’re simultaneously preparing to replace.

For solo builders, this means: don’t assume the built-in guardrails from OpenAI, Anthropic, or Google are enough. They’re a baseline, not a security strategy. The 71% of enterprise AI agents that are essentially chatbot wrappers have the same security problem at a smaller scale.

The bottom line

AI agents are useful — I’m not going back to doing everything manually. But the security practices haven’t caught up to the autonomy we’re giving these tools. Half of enterprises have already learned this the hard way. You don’t have to.

Start with an audit. Twenty minutes checking what your agents can access might save you from becoming a statistic.

If you’re just getting started with AI automation and want to build it right from the beginning, start here — I cover the setup process with security baked in from day one.