The Hidden Risk: Why 69% of Enterprises Are Exposing AI Agents
The rapid adoption of AI agents in the workplace has outpaced the security infrastructure designed to protect them. According to new research from VentureBeat, a staggering 69% of enterprises allow their AI agents to share credentials, creating a significant security vulnerability. This practice, where multiple agents operate using the same API key or service account, has become a primary target for cyberattackers who now view these shared keys as master keys to sensitive corporate data.
The core issue lies in the blast radius of a compromised credential. When an enterprise assigns a single API key to several different agents, a security breach in just one of those agents grants the attacker the combined permissions of all of them. Because the credentials are shared, there is often no clear forensic trail to determine which specific agent initiated a malicious action, making attribution nearly impossible. Furthermore, AI agents are uniquely susceptible to prompt injection attacks. If a malicious actor tricks an agent into revealing its configuration, the agent, having direct access to the raw API key, can be instructed to exfiltrate that key to an external domain. Once a static API key is leaked, it can be used to execute automated attacks at machine speed, causing significant damage before security teams can even detect the breach.
The industry is currently struggling with a lack of proper agent identity management. VentureBeat’s survey found that only 32% of companies assign each agent its own scoped, managed identity. Many organizations continue to rely on traditional identity models built for humans, ignoring the fact that AI agents operate continuously and can make thousands of requests per minute without responding to human-centric security checks like CAPTCHAs. The risk is not merely theoretical, as more than half of the surveyed enterprises, 54%, reported having already experienced an AI security incident or a near-miss. Larger organizations are at even higher risk, with incident rates climbing to 63% for companies with more than 1,000 employees.
To mitigate these risks, security experts recommend moving away from static, shared credentials toward a credential isolation model. Every AI agent should have its own identity, scoped to the minimum permissions required for its specific task. Implementing a local proxy layer between the agent and the API ensures the agent never sees the raw credential, as the proxy retrieves the key from a secure vault and injects it only when a request to an approved domain is made. For high-stakes operations, systems should require explicit human approval before the agent is granted permission to act. Finally, enterprises must log every access request with both the agent's identity and the delegating user's identity to ensure full accountability. As machine identities continue to proliferate, now outnumbering human identities by an 82-to-1 ratio in many organizations, the era of borrowed human credentials must end. Securing the agentic future requires treating AI-to-AI and AI-to-API communication with the same rigor as any other critical infrastructure.
