Shadow AI agents now work quietly inside company systems. The 2026 CSA survey and IBM report show the risks, from data leaks to rising breach costs.
Shadow AI agents are automated AI systems that employees install without IT oversight, security review, or formal approval. These agents run inside company systems and often slip past application inventories.
This type of agent carries out tasks on its own, opens connections through the Model Context Protocol, and stores credentials. The risk differs from that of passive SaaS apps, and older shadow IT frameworks were not designed to detect it.
Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025. That pace of adoption outstrips the governance most companies have built. As a result, a growing number of agents are running with no one keeping track of them.
The Latest Data Reveals a Visibility Gap
The Cloud Security Alliance (CSA) and Token Security survey, titled Autonomous but Not Controlled, describes the current state of shadow AI agents. Some 68% of organizations say they have high visibility into their AI agents. Yet 82% discovered at least one agent in the past year that their security or IT team did not know about. In addition, 65% experienced an agent-related security incident, with data exposure the most common impact.
Token Security funded the survey and sells identity security for AI agents. Readers should keep that in mind when weighing its conclusions. Only 14.4% of organizations deployed agents to production with full approval from their security or IT teams. Only 21% of respondents have a formal process for decommissioning agents.
Two other surveys show a similar pattern. Richmond Advisory Group, sponsored by Coalfire, surveyed 150 security leaders in North America in February. Of the incidents reported, 20% involved agents that created services without the operator’s knowledge, and 29% of respondents could not trace sub-agents. A survey of 235 security leaders found that 92% do not yet have full visibility into AI agent activity.
The IBM Cost of a Data Breach 2026 report puts a price on the problem. Ponemon Institute compiled it from 602 organizations that experienced a breach across 16 countries. Incidents involving shadow AI rose to 43%, up from 20% in 2025. The average cost of a shadow AI incident reached $5.39 million, compared with $4.63 million a year earlier.
The global average cost of a data breach is $4.99 million. IBM’s figures cover all unauthorized AI use, not just agents. Some 68% of organizations that suffered a breach lacked adequate AI governance, up from 63%. Among organizations with AI-related breaches, 92% lacked proper AI access controls.
Gartner estimates that Fortune 500 companies will use more than 150,000 AI agents on average by 2028, up from fewer than 15 in 2025. Today, only 13% of companies are confident their agent governance is adequate. An Okta survey in January 2026 found that only 27% of leaders consider their identity systems ready to manage non-human identities.
Recommended Control Measures
Gartner warns companies against blocking AI agents across the board, because employees typically turn to shadow AI instead. That approach can actually add risk. Companies need to regulate usage, not ban it.
According to Gartner, companies can start with four steps:
- Set a policy on who can create and share agents, and which connectors are allowed.
- Build a centralized agent inventory that also records shadow AI.
- Control what data agents can access.
- Provide training and build a community of practice around responsible AI use.
From an identity standpoint, each agent needs an owner, a defined scope of tasks, a credential lifecycle, and an audit trail tied to an accountable human. The CSA report recommends that companies treat agent creation as a governance event across all platforms. That way, every shadow AI agent is logged from the moment it is created, and its access can be revoked when it is no longer in use.
Get the latest information and other useful articles only at bacatulisan.com