Fortinet to acquire Virtue AI to bolster agentic AI security

Generative AI has already raised concerns about data leakage, unsanctioned use and inaccurate outputs. Agentic AI, the report says, is harder to secure and carries greater potential impact when breached. These systems can retrieve information, call application programming interfaces, trigger workflows, interact with applications and act on behalf of people, making an AI agent less like a chatbot and more like a digital employee with access credentials and permissions.

Traditional application security was designed for relatively predictable systems. Agents, however, dynamically decide which tools to invoke, what information to retrieve and what next step to take based on context. They can interact with multiple systems in a single workflow, drawing on customer data, internal knowledge bases, cloud services and business applications. In some cases, one agent may delegate tasks to another.

The report identifies several security concerns. Prompt injection can cause an agent to ignore its intended instructions or take unintended actions. Excessive permissions can turn a minor error into a significant incident. Connected tools can create indirect paths into critical systems, and sensitive data can be exposed through poorly governed retrieval or workflow execution. Because agents operate at machine speed, a flawed decision can escalate much faster than a traditional human-driven security event.

The biggest mistake enterprises can make, according to the report, is treating AI agents as just another application category. They are dynamic actors in the environment, with identities, permissions, behavioral patterns and connections to other systems. Static, point-in-time assessments such as pre-production testing, procurement reviews and periodic audits remain important but are insufficient for autonomous AI. An agent can behave differently tomorrow even if no one changes the underlying business process.

Continuous AI protection means validating an agent before deployment, monitoring it while it operates and reassessing it whenever its model, data, tools, permissions or role changes. It also means moving beyond simple guardrails at the model interface and gaining visibility into the agent’s decisions, tool calls and outcomes. The goal is not to stop enterprises from using agentic AI, but to enable them to deploy agents with confidence by establishing meaningful guardrails for their behavior.