54% of enterprises already had an AI agent incident

Everyone deployed AI agents first and figured out the controls later. Now there's a number on it. VentureBeat Research published findings today from 573 qualified respondents showing that 54% of enterprises had an AI agent security incident or near-miss in the past 12 months, and that more than a quarter don't learn what an agent costs until the invoice arrives. The AI agent governance gap isn't a prediction anymore — it's a retrofit bill coming due.

What actually happened

The research is five parallel surveys fielded in June 2026 across the agentic stack: orchestration (101 respondents), reliability and evals (157), security and identity (107), infrastructure and compute (107), and context layers/RAG (101). All respondents were at organizations with 100 or more employees.

Three numbers do the work. Security: 54% reported an incident or near-miss — 18% a confirmed incident, 36% a near-miss caught before harm. Cost: 27% exercise only reactive control over agent spend, with no per-agent budget or ceiling. Churn: in every one of the five control layers measured, 57–68% of enterprises plan to switch vendors or add new ones within 12 months, and roughly a third plan to move within the quarter.

Read those together and the story isn't that agents are dangerous. It's that companies shipped them knowingly ahead of the controls, and are now paying to bolt the controls on — which is always more expensive than building them in.

Why this matters for your business

You're not a 500-person enterprise, and that cuts both ways. You have fewer agents to govern. You also have no security team, no FinOps function, and nobody whose job is noticing that an agent quietly ran up $3,000 in tokens on a retry loop.

The three gaps map to three things you can do this week, none of which require a platform purchase.

Meter every agent separately. One API key per agent or per workflow, with a hard spend ceiling on each. Not for accounting — so a runaway loop fails instead of billing. That 27% figure is the cheapest problem on the list to fix and the most common one we find.

Give agents their own credentials. Not a shared key, not your personal token. Scoped, revocable, per-agent. When something goes wrong you want to kill one credential, not rotate everything.

Log the decision, not just the output. Which model, which tools, which inputs, what it did. The 36% who caught a near-miss caught it because something was visible. If your agent's work is a black box between prompt and result, you don't get a near-miss — you get an incident.

The vendor-churn number is the quiet warning. Two-thirds of these teams are about to swap governance tooling, meaning most of what got bought in 2025 didn't hold. Build controls into your own systems where you can, and treat any governance vendor as replaceable.

Sources: VentureBeat Research, VentureBeat.