AI Agent Security
You can't have AI security without API security. A prompt check can come back green while the APIs behind the agent are wide open, because every consequential thing an agent does is an API call.
How it actually works
AI agents are web apps built on APIs.
A chat interface connects through an API gateway, MCP and tool orchestration, LLM providers and backend services, all of it through APIs.
This is already happening
Three breaches, one pattern.
In each case the model did what it was built to do. The damage happened in the path behind it.
McDonald's and Paradox.ai
64M+ applicant records exposed
A hiring chatbot sat on an API that took a lead ID. Changing the ID, a BOLA flaw, returned other applicants' names, contact details and chat transcripts.
Figma MCP server
1 malicious parameter gave full server control
Remote code execution led to data theft and lateral movement, all of it looking like normal agent traffic.
OpenAI evaluation agent and Hugging Face
17,000+ recorded actions, none obviously malicious alone
The agent escalated privileges and moved laterally to a production database. Only the full sequence showed it.
The model worked. The path didn't.
Source: Anatomy of AI agent breaches whitepaper.
When agents take over, the edge goes blind
Agent traffic runs east-west, where the edge can't see.
A WAF or gateway inspects traffic that crosses the edge. Agents work inside, so one request at the edge can set off a chain of calls the edge never sees.
76.8%
say their agents chain tools and APIs together
12.5%
can consistently trace an agent's full path
17.5%
can stop an agent automatically
Salt Security, 2H 2026 State of Agentic AI and API Security (survey of 380+ professionals).
Where are you?
Tell us where you are in your AI journey. We'll help you get to the next level.
Pick the stage that sounds most like your program today. Every stage builds on the one before it, on the same graph.