The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents…

Across 107 enterprises, AI agents are already operating with real access to systems and data while the controls meant to contain them lag. VentureBeat reports that 54% of those organizations have already had a confirmed AI agent security incident or a near-miss. The pattern is not theoretical: agents are in production paths, and a majority of surveyed shops have already seen that exposure bite.

On the control plane, the gaps are structural. Only about a third of enterprises give every agent its own scoped identity. Most agents still share credentials rather than holding distinct, limited secrets. Only three in ten isolate their highest-risk agents. The security stack is overwhelmingly borrowed from model providers and hyperscalers, not purpose-built for agent identity, tool access, and blast-radius limits.

For engineers and builders, that combination is the operational risk. Shared credentials mean one compromised agent can inherit another’s privileges. Missing per-agent identities make audit, revoke, and least-privilege hard to enforce. Without isolation for the riskiest agents, a single bad tool call or prompt-driven action can reach systems that should never share a security boundary with a general-purpose agent.

Market context is that enterprises are adopting agents faster than they are building agent-native security. The default stack comes from the same model providers and cloud platforms that host the agents, which covers models and infrastructure well but leaves agent-specific controls thin: who the agent is, what it may touch, and whether high-risk agents run in a sealed environment.

Practical takeaway: treat agent access like service accounts under least privilege. Give each agent a scoped identity, stop credential sharing, and isolate highest-risk agents first. Watch whether teams move off shared secrets and provider-default controls toward agent-specific identity, isolation, and incident tracking—or keep running production agents on controls designed for chatbots and cloud APIs.