The AI Agents You Can’t See Can Still Create Risk

AI agent use is spreading rapidly among employees and developers through embedded enterprise software, no-code and low-code platforms, and custom-built automations. Employees often adopt these tools without waiting for formal approval, making discovery difficult for cybersecurity teams.

Organizations are already seeing significant levels of AI activity outside formal oversight. According to Gartner, unsanctioned AI use is already common, with 75% of organizations reporting unauthorized use of AI coding assistants and 50% reporting employee access to public SaaS GenAI agent platforms. This combination of strong business demand and limited visibility means security leaders cannot simply slow AI adoption. Instead, they must find ways to identify, inventory and monitor both sanctioned and unsanctioned AI agents.

The rapid growth of AI agents is forcing cybersecurity teams to rethink existing security programs. In order to do so, organizations must extend discovery capabilities beyond basic visibility, incorporate AI agent risks into posture management practices and update incident response processes to address rogue automations.

The challenge is becoming more urgent. Sixty-one percent of senior cybersecurity professionals have observed AI agent automation in existing approved enterprise software, while 59% suspect or have evidence of unsanctioned employee use. And according to Gartner polling, an average of 41% of stand-alone GenAI prototypes reach production, highlighting the need to pace security investments carefully and focus early efforts on data security and access management.

One of the biggest misconceptions about AI cybersecurity governance is that every AI agent requires the same level of oversight. When accessing risk, CISOs should start with a practical approach that prioritizes risk reduction based on two key factors: the sensitivity of the data an agent can access and its level of autonomy (the combination of agency and automation).

Some agents operate within defined enterprise applications, while others act independently, connect to multiple resources or generate business logic through large language models. As autonomy increases, so does the potential impact of misuse, compromise or unintended behavior. To gain meaningful oversight, it is advisable for cybersecurity leaders to create AI agent risk profiles, document each agent's permissions and scope, and categorize agents according to business risk. This approach helps organizations focus their attention where potential disruption is greatest. As this might not be fast enough, cybersecurity leaders must also adopt a “track back” inventory, for example by monitoring changes in utilization patterns for their sensitive APIs.