Shadow AI and agents: the new AI security risks in the enterprise
Artificial intelligence is already inside companies, whether it has been authorized or not. That is the recurring conclusion of the leading security reports of 2026, and it poses a new challenge: protecting not only the data, but also how AI tools and agents use it.
AI-driven attacks are rising and breaches cost more
IBM‘s Cost of a Data Breach Report 2026 puts the global average cost of a data breach at USD 4.99 million, up 12% from the previous year and an all-time high. The same report records a 56% increase in AI-driven attacks, led by deepfake impersonation and AI-powered malware (IBM, 2026).
“Shadow AI”: the AI the company can’t see
Shadow AI is the use of AI tools that the IT department has neither approved nor controls: pasting a contract into a public chatbot, summarizing customer data with a browser extension, or connecting an assistant to an email account. According to Verizon‘s Data Breach Investigations Report 2026, frequent use of AI tools by employees went from 15% to 45% in a single year (Verizon, 2026). That figure measures frequent AI use, not only unauthorized use, although Verizon links it to the rise of shadow AI, which is now the third most common non-malicious activity in data leaks.
The risk is already showing up in incidents. IBM’s 2025 report analyzed 600 organizations that suffered a breach. It found that one in five had been breached because of shadow AI and that a high level of shadow AI added USD 670,000 to the cost of a breach on average. In addition, 63% of those organizations had no AI governance policies to manage it or to prevent its unauthorized use (IBM, 2025; IBM X-Force, 2025).
Access control, the weakest link
IBM’s most revealing finding concerns access. 13% of organizations reported breaches of their AI models or applications, and of those, 97% acknowledged they lacked proper AI access controls (IBM, 2025). In other words, the problem is usually not a sophisticated attack on the model, but who can use it, with what data and with what permissions.
AI agents raise the bar
With AI agents, which plan, remember, use tools and act with delegated authority, the risk changes in scale: an agent with excessive permissions can read, modify or send information without anyone reviewing it. That is why OWASP‘s generative AI security project published the Top 10 for Agentic Applications for 2026 in December 2025, a risk catalog specific to agentic applications that serves as a reference for designing and auditing them (OWASP, 2025).
Five controls to get started
- Inventory: identify which AI tools are actually used in the company, not just the approved ones.
- Approved alternatives: banning without offering options pushes people toward more shadow AI; provide corporate tools that keep data protected.
- Least privilege: every user, integration and agent should access only the data and actions it needs.
- Human oversight: an agent’s sensitive actions (payments, sending data, deletions) should require approval.
- Training and a clear policy: everyone should know which data they can share with an AI tool and which they can’t.
Security is not a brake on AI; it is the condition for scaling it with confidence. At Wasivia, we design AI and agent solutions with governance, access control and traceability built in from the start. Let’s talk about your case.
