Ghost AI agents: RSA Agent ID in action

A mid-size global bank claimed it had no AI agents in production. Internal policy banned them outright. The audit found more than 4,000 agents running.

That is the story Jim Taylor, President of Product and Strategy at RSA, shared at the Agent ID launch during the AI Conference in San Francisco on September 29. RSA Agent ID is an identity platform for AI agents: discover, secure and govern them, with a gateway into AI and MCP ecosystems.

Numbers that are hard to process

Gartner expects roughly 150,000 AI agents in a typical Fortune 500 enterprise by 2028, up from fewer than 15 in 2025. Only 13% of organizations believe they have the right governance in place. Per IBM, an incident involving shadow AI costs on average $670,000 more than a conventional one.

The prompt that took a bank offline

The most striking story involves no attacker at all. A support employee asked an agent to pull "all the data" from Salesforce to build charts. The agent downloaded the entire database, Salesforce flagged abnormal traffic and shut the instance down believing it was stopping a denial-of-service attack. One poorly worded prompt brought the company to a standstill.

The root problem is identity

Identity is already a hard problem for humans and service accounts. An AI agent is neither: it accumulates permissions, nobody revokes anything, and a customer prompt can push it outside its original scope.

Emerging solutions — RSA Agent ID or the open source Nvidia platform released the same week — all do the same three things: discover the agents, assign each one a human owner, log and control every call.

If you run agents in production, the inventory audit comes before the tooling. You cannot secure what you don't know exists.