Enterprise AI agent deployments doubled inside most organisations in the first half of 2026. The security incident rate followed. According to DigiCert's survey of 1,001 IT and cybersecurity decision-makers, published July 7, 78% reported AI-related security incidents in the past six months — with half of those incidents tied directly to an unauthorised or misconfigured AI agent.
What Actually Happened
Three converging sources published findings this month that tell the same story from different angles.
DigiCert's research found that while 75% of organisations deployed four or more AI-powered systems in the past six months, almost half lack centralised visibility into AI system activity. Only 40% have assigned unique digital identities to at least some of their agents. And 47% cannot trace AI decisions back to the models and source data that produced them — a gap that makes post-incident investigation nearly impossible.
Gravitee's State of AI Agent Security 2026 report added a structural dimension: 80.9% of technical teams have pushed agents into active testing or production, but only 14.4% did so with full security and IT approval. Mean monitoring coverage across deployed agents sits at 52%, meaning roughly half are running with no consistent logging or oversight.
The Sophos AI Security 2026 Report, published July 22, identified a distinct but related attack vector: AI identities — the credentials, service accounts, and trust relationships agents use to operate — are being targeted specifically. Attackers are exploiting agent credentials not to compromise the agent directly, but to move laterally inside enterprise networks using the access the agent was legitimately granted.
eSecurity Planet's weekly security roundup, published July 28, named AI agents, trust abuse, and breaches as the defining themes of enterprise security this week — a characterisation that has recurred across multiple consecutive weekly roundups.
Why This Is Harder Than It Looks
The intuitive response is to slow deployment until governance catches up. The data suggests this is not what is happening, and understanding why matters more than repeating the recommendation.
81% of technical respondents in the Gravitee survey reported feeling pressure to deploy agents quickly, even when security was not ready. That pressure is not irrational. AI agents that automate high-friction internal processes deliver measurable productivity gains quickly. Organisations that wait for a complete governance model before deploying will find themselves explaining the delay while their peers report output improvements.
The deeper problem is that existing security infrastructure was not built for this class of entity. Human identity governance systems — PAM tools, identity lifecycle management, SIEM platforms — assume the identity behind an action is a person. AI agents are not persons: they authenticate continuously, execute at machine speed, call external APIs, and produce outputs without pausing for human review. Most enterprise security stacks have no mechanism to capture, monitor, or alert on agent-to-system interactions at the granularity required.
The Sophos finding on AI identity as an attack vector points to a specific gap: agents are commonly provisioned with broad access permissions because scoping them precisely requires more upfront engineering than most teams invest. That over-provisioning is now an exploitable surface. Least-privilege access — a well-established principle — has simply not been applied to agent fleets yet.
The Enterprise Lens
If your business uses any AI-powered tool that connects to internal systems — a customer service bot that reads order data, a workflow automation that processes invoices, a sales assistant that accesses your CRM — you have AI agents running in your environment. The question is not whether you have exposure; it is whether you can see it.
The starting audit has three questions: What AI agents are currently running and what systems can they access? What authentication credentials do those agents use, and are those credentials unique per agent or shared? Is there a log of what those agents have done in the last 30 days?
Most organisations that have run this audit report two consistent findings: agents with significantly broader access permissions than their use case requires, and insufficient logging to reconstruct agent activity after the fact. Both are fixable without rebuilding your security stack. The fixes involve access scope review and logging configuration, not new infrastructure.
The DigiCert finding that 47% of organisations cannot trace AI decisions back to their source models and data is the clearest operational risk. When a client asks how a specific outcome was reached, or when a process fails and root cause analysis is required, the inability to reconstruct the reasoning chain is a liability — legal, operational, and reputational.
What to Watch
- Whether enterprise AI platform providers — Microsoft, AWS, Google — move to embed agent identity governance and audit logging directly into their orchestration layers. This would reduce implementation cost significantly versus building controls at the application layer.
- How the EU AI Act's high-risk system classification (enforcement began August 2) applies to agents that contribute to decisions affecting individuals. Organisations in regulated sectors may discover compliance obligations around their agent deployments that were not anticipated when those agents were built.
- The emergence of dedicated AI agent security tooling as a vendor category. Several early products have appeared in the past quarter; consolidation and enterprise adoption patterns will indicate whether the market is solving this problem at scale or fragmenting it further.
Sources
- DigiCert: 78% of Enterprises Report AI Security Incidents
- State of AI Agent Security 2026 Report — Gravitee
- AI Agents, Trust Abuse, and Breaches Define Cybersecurity News this Week — eSecurity Planet
- The Enforcement Gap: 88% of Enterprises Reported AI Agent Security Incidents — VentureBeat
- The Security Risks of AI Agents in the Enterprise — Security Boulevard