Agentic AI Security Gap: Firms Expect Incidents, Spend Just 6%

Enterprises are deploying autonomous AI agents faster than they are learning to defend against them. Two new industry surveys point to the same tension. Security leaders overwhelmingly expect AI agents to cause serious trouble, yet their budgets, governance and tooling have not caught up. Vendors, meanwhile, are moving quickly to sell into that gap.

The numbers behind the anxiety

The starkest figures come from a 2026 Agentic AI Security Report. It surveyed 300 enterprise leaders worldwide across security, fraud, identity and AI functions. Ninety-seven percent of respondents said they expect a material security or fraud incident driven by AI agents within the next 12 months, and 49% expect one within six months. 2

Spending tells a different story. On average, organizations devote only 6% of their security budgets to AI-agent risk. One in ten does not track that risk as a separate category at all, and more than half have no formal governance controls for AI agents. 2 More than 70% of security teams also doubt that their current tools will scale as AI-driven attacks evolve. The report frames this as a "detection illusion," in which defenses that seem adequate today may not hold up. 2

Darktrace's State of AI Cybersecurity 2026 points in the same direction from a different angle. In that survey, 92% of security professionals said they are concerned about the use of AI agents across the workforce and what it means for security. 1 Darktrace locates the core problem in permissions. Agents in production often act across multiple systems with broad access to sensitive data, business-critical applications, tokens, APIs, and even IT and security tools. 1

The two reports measure slightly different things. One asks about expected incidents and the other about general concern. Even so, they arrive at near-unanimity: practitioners see agentic AI as a real and near-term risk.

Where the worry concentrates

Darktrace's data also covers generative AI more broadly, especially what employees type into prompts. Among respondents:

• 61% named sensitive data exposure as their top concern.
• 56% pointed to data security and policy violations.
• 51% cited misuse or abuse of AI tools. 1

These worries overlap with the agent problem. An agent with wide permissions that is fed sensitive context is essentially a prompt-leakage risk that can also take actions. Darktrace also notes that attackers are using AI as bait, for example with fake AI software installers that deliver malware. 1 The threat runs both ways: AI agents can become an internal liability, and the AI brand itself can become an external lure.

Vendors move into the gap

The supply side is responding. At its Cyber Summit in Paris on October 1, 2026, Thales announced five cybersecurity products and solutions. 3 They include:

• A hardware security module built for post-quantum protection.
• An AI-enabled data security posture management (DSPM) product.
• Software protection designed to resist AI-powered attacks.
• An expanded Google Cloud partnership aimed at securing agentic AI workflows.
• A new global approach centered on real-time threat detection and response. 3

The Google Cloud tie-up is the most directly relevant to the survey findings, because it targets the workflows where agents operate. The DSPM product addresses the data-exposure concern that topped Darktrace's list. Thales frames the launch as a response to increasingly automated and AI-driven threats. 3

Reading the gap

Some caution is warranted. Both surveys come from security vendors, who benefit when buyers feel alarmed. A figure like 97% expecting an incident within a year partly reflects how the question was asked. The survey defined the threshold as "material," but answers depend on how each respondent interprets that word.

Still, the more telling data points are the structural ones, not the predictions. These include the 6% budget share, the lack of governance at more than half of organizations, and the fact that one in ten does not track the risk at all. 2 These are concrete descriptions of current practice, and they match Darktrace's account of agents operating with sweeping, poorly bounded access. 1

This looks like a familiar pattern from earlier technology shifts such as cloud and SaaS. Adoption runs ahead of governance, and security spending follows only after a visible incident forces the issue. The difference this time may be speed. Agents act autonomously and at machine pace, so the window between a misconfigured permission and real damage could be shorter than in past cycles.

The practical lesson for security leaders is less about buying new products and more about basic discipline. That means inventorying agents, scoping their permissions tightly, tracking agent risk as its own category and setting governance rules before deployment, not after. Vendors like Thales are betting that budgets will eventually follow the fear. Survey respondents already expect an incident, and the open question is whether their organizations act before one happens.