Why put AI security on a SASE scorecard?
A SASE service already sits between users and the internet. When staff use AI tools, or when AI agents call tools and services, that traffic passes through the same inspection point. Adding AI controls to the platform you already run can mean one policy for web, SaaS and AI traffic instead of a separate product. That is why AI security inside the platform carries 15% of our total.
Which controls should you compare?
Five areas cover most needs. Discovery: can the platform find which AI tools and agents are in use. Prompt and data controls: can it inspect what is sent and redact or block sensitive data. Agent and MCP controls: can it see and govern AI agents and the Model Context Protocol servers they call. Runtime protection: can it stop prompt injection and jailbreak attempts against your own AI apps. Enforcement: does it block inline, or only alert.
What do the vendors list?
Cato states AI security for users, applications and agents with one engine and one console, prompt controls with Monitor, Anonymize and Block actions, and mapping of agents, MCP servers and skills. Netskope lists an AI product family that includes an Agentic Broker for MCP traffic and AI Guardrails. Zscaler lists AI asset discovery including MCP servers and an AI Gateway with MCP and A2A brokers. Fortinet describes FortiAIGate checks on prompts, outputs and agent tool calls, including MCP tool traffic, as a self-hosted product. Palo Alto Networks lists AI Access Security and Prisma AIRS; MCP is not named on the pages we read. Cisco describes AI Access in Secure Access and agent controls; MCP is not named on the pages we read.
Source: Cato AI security
Source: Netskope AI products
Source: Zscaler AI security
Source: Fortinet FortiAIGate
Source: Palo Alto Networks AI Access Security
Source: Cisco Secure Access AI Access
Does it matter where the control runs?
Yes, for overhead. A control inside the SASE cloud uses the same policy and logs as the rest of your traffic. A control delivered as a separate gateway or appliance can still work well, but it is another system to deploy and keep in step.
How should you test AI controls?
Run real prompts that contain test data patterns your DLP should catch. Try an unsanctioned AI tool and an approved one. If you run agents, connect one to a test MCP server and check whether the platform sees and can stop a tool call. Record whether each test was blocked, redacted or only logged.
How does this site score it?
AI security inside the platform is scored from what each vendor lists across the five areas above, with credit for naming MCP and agent controls and for inline blocking. Each profile shows the current score and its one-line reason. The weighting lab has a must-have filter for vendors that name MCP or agent tool-call controls.