Enterprises Fail to Isolate High-Risk AI Agents

A new VentureBeat Pulse Research study reveals a significant vulnerability in enterprise AI agent security. While 65% of organizations enforce scoped permissions at runtime for their AI agents, a mere 18% isolate their highest-risk agents. This finding highlights a growing risk as AI agent autonomy scales.

The survey of 116 enterprises found that 53% have AI agents in production, with an additional 27% piloting them. Agent-specific security incidents or near-misses have already occurred in 53% of organizations, contributing to a declining confidence in agent security. Credential sharing persists across nearly two-thirds of agent fleets.

The core issue identified is a containment gap. Enterprises have implemented controls for monitoring and permissions but lack robust mechanisms to limit damage when these fail. Only 21% of agents in production are isolated, and just 8% pair isolation with permission enforcement, an order reversed from a defense-in-depth strategy.

Confidence in security has slipped, with 30% of enterprises believing AI-armed attackers are ahead of their defenses. Despite high satisfaction with current tools, 74% plan to adopt, add, or replace agent security tooling within twelve months, indicating a strong demand for improved security measures.