54% of Enterprises Have Already Had an AI Agent Security Incident

More than half of surveyed enterprises — 54% — have already experienced a confirmed agent security incident or near-miss, according to new research from VentureBeat's AI vertical. Yet only about a third of companies provide each agent with its own scoped identity, leaving credential-sharing as the dominant deployment pattern.

The blast radius problem

When a single service account is shared across dozens of agents, a compromise anywhere in the graph gives an attacker access to everything the graph could touch. In one representative incident, a jailbroken customer-service agent gained read access to the entire company's CRM because it inherited the same role as the internal analytics bot.

Why identity lags

Most identity providers weren't built for machine identities that spawn, act, and terminate within minutes. Vendors like Okta, Microsoft Entra, and newer startups (Sesame, Vault) are racing to ship per-agent identity primitives that scale to those numbers.

The short-term fix

Security teams are increasingly forcing agents through API gateway proxies that enforce per-call scoping, rate limits, and audit logs. It's a stopgap, but for now it's the most concrete lever.

Source: VentureBeat