RSA Launches Agent ID to Find and Police Shadow AI Agents in the Enterprise

RSA unveiled Agent ID at The AI Conference in San Francisco, an identity security platform built to track AI agents operating inside large organizations. The product targets finance, government, healthcare, and critical infrastructure, sectors where an autonomous agent taking the wrong action carries compliance and safety consequences. RSA's framing is blunt: agents are proliferating faster than security teams can inventory them, and most enterprises do not know how many are running or who authorized them.

The company says the problem is not hypothetical. In a discussion tied to the launch, RSA's Jim Taylor described an incident in which a single malformed prompt took down a company's Salesforce environment. The scale of the exposure is the harder number to fix: RSA estimates roughly 4,000 shadow AI agents can hide inside a typical large enterprise, spun up by employees and teams without security review. That figure is the market RSA is selling into.

Agent ID ships in three parts. Discover locates both sanctioned and unsanctioned agents, along with the Model Context Protocol servers they connect through. Secure is an inline gateway that inspects every tool call an agent makes and checks it against organizational policy before the action executes. Govern maps the resulting evidence to ten regulatory frameworks, giving compliance teams an audit trail rather than a log file.

The inline gateway is the piece that separates Agent ID from conventional SaaS discovery tools. Finding an agent tells you it exists; intercepting its tool calls is what prevents an agent from reading a database it should not touch or firing an API request that moves money. MCP servers matter here because they are the connective tissue between models and enterprise systems, and they are frequently stood up by developers without a security review.

Discover and Secure are scheduled to ship on November 16, 2026. RSA has not published pricing or said how Govern will be licensed, and the gap between now and general availability leaves room for competitors in agentic identity to move first. The open question is enforcement: whether buyers want a gateway that can block an agent mid-task, or a monitoring layer that flags problems after the fact.