54% of enterprises already had an AI agent incident. Most still share credentials.

107 enterprises surveyed.

54% already had a confirmed agent security incident or near-miss.

18% confirmed. 36% caught it before it broke.

Here's what should keep you up tonight:

69% of those enterprises share credentials across their agent fleet.

Only 32% give each agent its own scoped identity.

Only 30% sandbox their highest-risk agents.

The blast radius math is simple. When an agent shares credentials with other agents, one compromised agent compromises all of them. Forensics can't tell which agent did what. The non-human identity problem is the single largest unfinished piece of enterprise agent security.

The false comfort is worse than the gap. Satisfaction with current tooling: 4.2 out of 5. Most are using OpenAI's guardrails (51%), Google's controls (36%), or Microsoft's Azure (35%). Purpose-built agent security barely registers. Palo Alto, CrowdStrike, Zenity — all in the low single digits.

82% name a provider-native control as their primary security layer.

Yet only 35% believe their defenses are ahead of AI-enabled attackers. 21% say attackers are ahead. A clear majority plan to rip out their current stack within 12 months.

They're satisfied with controls they're simultaneously preparing to replace. That is the definition of false comfort.

The EU AI Act's high-risk obligations hit August 2. If your agents touch critical infrastructure, financial systems, or employment decisions, Article 12 requires logs that reconstruct what the system did and why. Shared credentials make that impossible.

Audit your agent identity posture today. If any agent in your fleet shares a credential with another agent, you are one compromised prompt away from a breach you cannot attribute, cannot contain, and cannot explain to a regulator.

SOURCE: https://venturebeat.com/ai/the-agent-security-gap-54-of-enterprises-have-already-had-an-ai-agent-incident-and-most-still-let-agents-share-credentials
VERIFIED: VentureBeat Pulse Research (July 16, 2026), 107 enterprise respondents
SIGNAL: Agent adoption is outrunning agent security. Identity and isolation — the controls that matter most when something fails — are the ones enterprises have built least.