Verifiable Execution Isolation Layer - VEIL

AI agents are increasingly capable of reading emails, documents, databases and other external information, and then taking actions through tools and APIs. This creates a security problem: an agent can be influenced by untrusted content, such as a malicious email or document, and turn that influence into a dangerous tool call. The security risk is therefore not limited to the prompt itself; it exists at the point where an agent attempts to execute an action. VEIL — Verifiable Execution Isolation Layer — addresses this problem by placing a runtime security boundary between an AI agent and the tools it can use. Instead of trusting the agent to make every security decision, VEIL intercepts proposed actions before execution and verifies them against capabilities, policies, data sensitivity and behavioral context. The system follows: AI Agent → VEIL Gateway → Security Core → Tool / Sandbox The gateway normalizes incoming actions and passes them through deterministic security controls. VEIL evaluates what capability the agent is requesting, what resource it is accessing, what type of data is involved, and whether the action forms part of a suspicious sequence. When an action requires deeper investigation, VEIL can use specialized security agents. VEIL-Counter identifies how a proposed action could be abused, while VEIL-Judge evaluates the threat hypothesis. The Decision Engine then enforces an outcome such as ALLOW, WARN, BLOCK or REVOKE. Our demonstration uses a controlled Doomsday Corp sandbox. A malicious input attempts to influence an agent into reading sensitive customer information and transferring it externally. Without a runtime boundary, the dangerous action can reach the tool. With VEIL, the proposed action is intercepted, analyzed and blocked before tool execution. The target users are organizations building or deploying AI agents that interact with sensitive data, enterprise tools, APIs and external systems. VEIL is designed to be model-agnostic, allowing the security boundary to remain independent of the underlying AI model. What makes VEIL different is its focus on execution rather than simply detecting malicious prompts. Even when an agent is compromised or manipulated, the action still has to cross VEIL before reaching a protected tool.