An employee connects an AI agent to the company customer relationship management (CRM) platform to save time. The agent can now access customer records and interact with other business systems, but the security team may have no idea the agent exists or what it might do with that access.
This is the risk of unapproved tools and autonomous agents operating inside enterprise environments, known as shadow AI, potentially exposing sensitive customer data or taking actions without adequate oversight.
Recent research indicates the threat is already producing operational consequences. In an October 13 report, Salt Security found that 49.8 percent of more than 380 security and technology professionals had confirmed or suspected that an AI agent in their organization took an action during the previous 12 months that was unintended, unexpected or unauthorized.
Among those reporting such incidents, more than half cited an unauthorized external API call or transaction, while more than 40 percent reported an agent modifying or deleting data. The research also found that only 12.5 percent of respondents could consistently trace an agent's complete activity from its initial instruction through connected tools and APIs to the outcome.
A September survey from EY found a similar gap between AI adoption and oversight. Although 98 percent of senior AI executives said that their organizations had formal AI governance policies, 26 percent of those using agentic AI admitted they could not detect unauthorized agents operating internally. Meanwhile, 36 percent reported an AI incident or failure that had caused material harm, including data loss, financial damage, operational disruption or reputational damage.
The implications increasingly reach beyond IT security and into the scope of customer experience leaders. An unmanaged agent with access to CRM records, payment information or customer conversations could expose personal data, change account details or initiate actions that customers never authorized.
Why Shadow AI Is Harder to Detect Than Shadow IT
Speaking to CX Today during Cybersecurity Awareness Month, Ankita Gupta, Co-Founder and CEO of Akto, an agentic AI security platform for enterprise security teams, described shadow AI as one of the most pressing challenges facing enterprise security teams.
“Shadow AI is the number one cyber AI risk today and it is extremely difficult to track for multiple reasons.”
Shadow AI is harder to track than the shadow IT that happens when employees use software, services and devices without formal approval.
Employees can access AI through a range of possible entry points, making it difficult for security teams to maintain a reliable inventory of what is running across the business.
As Gupta explained, “they are using AI tools in browsers, they are using AI tools in the cloud, they are using SaaS agents in something like a Microsoft Copilot or Salesforce Agent Post. They're using desktop agents; they're using IDEs which have AI features.”
Traditional IT discovery processes can miss AI capabilities embedded within existing software or introduced through an employee's personal workflow.
“Shadow IT in twenty years hasn't changed a lot,” Gupta said. “But now in the last two years, the way AI is changing so quickly and moving so fast, that speed is very hard to match for cyber teams as well.”
For large enterprises, the challenge multiplies across departments and employee populations. A single organization might have thousands of people experimenting with different AI applications, each with its own integrations, access requirements and potential weaknesses.
When An AI Agent Can Act on Customer Data
The security implications become more serious when employees move from using AI to generate content or answer questions to deploying agents that can perform tasks.
“Agents change the game a lot,” Gupta said. “They are autonomous in nature, they have access to lots of sensitive data.”
