AI Agents Outpace Security Controls as Organizations Grant Broad Permissions
A new survey reveals that 70% of organizations have granted AI agents decision-making and action-taking authority, leading to widespread security issues.
Organizations are rapidly delegating critical tasks to AI agents, with a significant majority now allowing these systems to make decisions and execute actions that previously required human oversight. A survey of 1,100 IT and security decision-makers found that 70% of companies have granted AI agents this level of autonomy, a stark increase from previous years. This shift reflects the growing capabilities of AI but also introduces substantial security challenges.
The consequences of this rapid delegation are already apparent. Three-quarters of organizations reported experiencing at least one AI-related security incident or warning sign within the past year. Common issues include AI agents accessing unauthorized systems, handling sensitive data without proper permissions, or initiating unapproved actions. Approximately a quarter of respondents encountered each of these specific problems, underscoring the immediate risks associated with unchecked AI agent activity.
Unlike simple chatbots, these AI agents are designed to act autonomously. They can interact with other systems, modify configurations, and execute multi-step processes without direct human intervention. Alarmingly, about 30% of organizations would permit an AI agent to conduct security investigations or alter production environments without any human review, highlighting a concerning level of trust in autonomous systems.
While many security leaders express confidence in their visibility into AI operations, a closer examination reveals a significant gap. 84% claim to know how AI systems access data and systems, but this figure drops to 49% when asked about the ability to audit every single action an agent takes. This lack of granular auditability becomes critical when incidents occur, hindering effective investigation and response.
Compounding these visibility issues, 62% of respondents admitted to deploying AI tools before establishing adequate security safeguards. This reactive approach leaves organizations vulnerable, as the focus is often on implementing controls after a problem has emerged rather than proactively preventing it. The survey also highlighted a paradox in security priorities: while most agree that limiting AI access is paramount, defining explicit agent permissions ranked last among chosen security controls.
Experts suggest that the difficulty in precisely defining AI agent permissions and the desire to avoid disrupting AI workflows contribute to this security gap. Raghu Nandakumara, VP of industry strategy at Illumio, notes that organizations often delay implementing strict controls, assuming they will address them once the agent's behavior is better understood – a point that rarely arrives. This friction between usability and security is a key driver of the current challenges.
The question of accountability for AI agent mistakes is also a significant concern. Legal and regulatory frameworks consistently place responsibility on the organization that deploys the AI, not the vendor. Within an organization, the business leader who approved the AI use case should ideally own the outcome. However, security teams often become the default owners when accountability is unclear, with CISOs bearing the responsibility for guardrails and agent access controls.
Regulatory bodies are beginning to address this, with financial services sectors leading the way in demanding named accountability for AI outcomes. Nandakumara advises organizations to start by securing the most critical AI agents first, focusing on limiting their access to the most sensitive data and systems. Implementing even a few essential controls on high-risk agents can provide immediate security benefits, far outweighing the pursuit of a perfect, comprehensive policy that may never be fully realized.