AI security startup Zenity has unveiled what it describes as the industry’s first security platform designed specifically for autonomous AI agents, introducing a decision-centric security architecture aimed at governing AI actions before they are executed across enterprise systems.
The expanded platform addresses a growing challenge facing organisations as artificial intelligence evolves from assisting employees with isolated tasks to independently executing complex, multi-step business processes. Unlike traditional AI applications that respond to individual prompts, autonomous AI agents increasingly perform long-running workflows, write software code, access enterprise applications, handle sensitive information and make operational decisions with limited human intervention.
As enterprises deploy these systems across software development, customer service and business operations, cybersecurity experts have warned that conventional security models—which typically detect threats after an action has occurred—may be insufficient for AI agents capable of acting independently over extended periods.
Zenity said its latest platform introduces what it calls “security at the decision layer,” evaluating every AI decision before it becomes an enterprise action rather than relying primarily on post-event monitoring and incident response.
The platform expands Zenity’s existing AI security capabilities with two new components—Exposure Management and Runtime Boundaries—designed to continuously assess AI risk and enforce enterprise policies in real time.
According to the company, the architecture evaluates an AI agent’s intent, identity, requested action, accessed data, available tools, execution history and organisational policies before determining whether an action should proceed, be blocked or be terminated.
The approach is intended to address risks associated with so-called long-horizon AI agents, which can autonomously perform tasks spanning hours or even days. While each individual action may appear legitimate, cumulative decisions across an extended workflow can create security vulnerabilities as an agent’s context, permissions and interactions evolve over time.
Rather than monitoring isolated AI prompts, Zenity’s platform analyses decision sequences across entire workflows, enabling organisations to identify risks before they result in unintended business actions.
The company’s security architecture is organised around three integrated functions—Surface, Enforce and Protect.
The Surface layer continuously discovers AI agents deployed across an organisation, identifies exploitable attack paths and prioritises potential exposure risks.
The Enforce layer applies Runtime Boundaries to evaluate every AI action in real time. These policy-driven controls allow enterprises to define what AI agents are permitted to do, from accessing sensitive corporate data to executing privileged administrative tasks or interacting with external systems.
The Protect layer combines AI-powered Digital Forensics and Incident Response (DFIR) with so-called Guardian Agents that learn from previous investigations to refine future security policies and improve threat detection over time.
According to Zenity, the platform creates a continuous security feedback loop in which exposure analysis informs runtime enforcement, while incident investigations strengthen future policy decisions.
The Runtime Boundaries engine supports a broad range of enterprise AI platforms, including Anthropic’s Claude Code, Cursor, Microsoft Copilot, Salesforce Agentforce, OpenAI’s ChatGPT Enterprise, Amazon Bedrock, Azure AI Foundry and custom-built AI agents.
The company said the technology can be used to prevent sensitive data exposure, govern AI coding assistants, restrict privileged system actions and manage Model Context Protocol (MCP) servers across enterprise AI deployments.
“Every major technology shift has forced security to evolve,” said Tomer Teller, Vice President of Product at Zenity.
“Autonomous AI introduces the next evolution: the decision itself,” Teller said. “The future of AI security won’t be defined by what an AI agent already did. It will be defined by what AI is allowed to do before it acts.”
Zenity said Exposure Risk capabilities identify the attack paths most likely to be exploited and feed those insights into Runtime Boundaries before AI agents execute actions. Following an event, the platform’s AI-powered Digital Forensics and Incident Response capabilities reconstruct decision chains to identify how actions occurred and automatically improve future governance policies.
The launch comes as enterprises increasingly move beyond generative AI experimentation toward deploying autonomous AI agents capable of independently executing business processes. While these systems promise productivity gains, they also introduce new security challenges as AI gains access to sensitive enterprise systems, proprietary data and operational workflows.
Industry analysts have increasingly identified governance and runtime control as emerging priorities for organisations adopting agentic AI, particularly as businesses deploy multiple AI agents capable of interacting with one another and operating without continuous human oversight.
Zenity said its latest platform builds on years of securing AI deployments within Fortune 500 organisations and reflects a broader shift in enterprise cybersecurity, where protecting AI systems increasingly requires governing decisions before execution rather than responding after actions have already occurred.
