Salt unveils platform to secure rising AI agent stacks

Salt Security has launched what it calls an agentic security platform designed to map and monitor how AI agents connect to enterprise systems through large language models, Model Context Protocol (MCP) servers and application programming interfaces (APIs).

Salt positions the Agentic Security Platform as a response to the growing use of AI agents in business processes. Many deployments now extend beyond chat-style interactions to trigger actions in internal applications and access data through interfaces originally designed for human-driven software.

Salt describes this as an "agentic stack" with three elements: LLMs as the reasoning layer, MCP servers as the execution layer that brokers tool use, and APIs as the interfaces that translate agent intent into actions in enterprise environments.

These connections form what Salt calls an "Agentic Security Graph". The concept treats the links between models, MCP servers and APIs as security context for understanding what an agent can access and what operations it can perform.

API security has long been a concern for security teams because APIs often expose business logic and sensitive data. Agent-driven activity increases the focus on APIs because agents can issue large volumes of requests and chain actions across systems. Salt argues that the operational risk lies less in an agent's text outputs and more in the transactions it can initiate when it has access to tools and services.

Salt says the platform provides a unified way to discover, visualise, govern and protect connections across LLMs, MCP servers and APIs, contrasting it with tools that focus on individual parts of the AI environment.

Posture and response

The release introduces two features: Agentic Security Posture Management and Agentic Detection and Response, abbreviated as AG-SPM and AG-DR.

Salt says AG-SPM provides continuous discovery and governance of LLM connectivity, agents, MCP servers, APIs and their relationships, including visibility into which systems are connected and how those connections are configured.

AG-DR focuses on runtime monitoring, with real-time detection of abuse, misuse and anomalous behaviour across LLM connectivity, agent-driven activity, MCP servers and APIs.

Salt describes the two features as covering "code to runtime", aligning the message with a broader shift toward earlier visibility into misconfigurations and clearer monitoring once systems are live.

"Most AI security solutions focus on prompts and models," said Roey Eliyahu, CEO and Co-Founder of Salt Security. "But the real enterprise risk is not just in what an agent can say. It is in what an agent can do through MCP servers and APIs. These systems connect agents to data, workflows and enterprise services. That is what we call the Agentic Security Graph, and Salt's Agentic Security Platform is designed to expose and secure it."

Customer feedback

Salt says early customers have used the platform as they expanded AI agent deployments, citing the growing challenge of tracking what agents interact with as the number of connected systems increases.

"As we deploy more AI agents across our organization, the complexity of the systems they interact with has increased dramatically and is challenging to manage," said a CISO at a large technology company. "Salt is uniquely positioned to secure this new environment because every agent interaction ultimately runs through APIs. The Agentic Security Platform gives us the visibility and protection we need to confidently scale AI across the business."

Enterprise context

The announcement comes as more vendors attempt to define security categories around AI agents and tool use. Security teams have already had to adapt to model and prompt risks, including data leakage and unsafe outputs. Agentic deployments add a different set of concerns because they connect models to systems that can approve payments, alter records, manage infrastructure or trigger operational workflows.

Salt's approach centres on mapping relationships rather than only monitoring prompts and responses, reflecting a broader need for inventory and governance as organisations connect multiple models to multiple internal tools and external services.

Salt also links the platform to growth in "autonomous interactions" across enterprise environments. Increased automation can raise the impact of configuration errors and credential misuse, since an agent may execute steps at machine speed and at scale once it has access.

Alongside the product launch, Salt is offering a limited number of "Agentic Security Graph Discovery Sessions" during the RSA Conference. Salt positions the sessions as a way for organisations to see how LLMs, MCP servers and APIs connect across their environments.