Your AI Agents Read Everything. Now You Can See What’s Trying to Manipulate Them.

The biggest risk for your organization may be something else than user prompts.

AI agents act on instructions you may not see. Every tool description they load, every response  a tool returns, every document they read is a potential instruction. Attackers know this. A poisoned tool definition or a manipulated tool result can quietly redirect an agent to leak data, escalate privileges, or take actions your policy never allowed. The industry calls this prompt injection, and it is consistently ranked the top security risk for LLM applications.

Most security teams can’t see it happening. Agentic traffic flows through Model Context Protocol (MCP) connections between agents and tools. Traditional controls don’t inspect it. If an instruction hides inside a tool result, nobody knows.

We are changing that.

Introducing Prompt Attack Detection

Workforce AI Security is opening Early Availability for Prompt Attack Detection. Prompt Attack Detection runs Check Point AI detection engines across agentic traffic in your organization. It analyzes:

  • Tool definitions, when agents load them, catching poisoned instructions before an agent ever acts on them
  • Tool call inputs and outputs, catching injection attempts that arrive through the data your agents consume

Every detection appears in your Agentic Events log with the detector name, the sensitivity level that triggered it, and clear attribution to input or output. You see exactly what was flagged, where it came from, and which agent and tool were involved.

Detect first. Enforce with confidence.

Early Availability runs in detect-only mode. Nothing is blocked. No workflows break. Your team gets full visibility into prompt attack signals across real agentic traffic, with zero operational risk.

Every detection carries feedback controls right in the event drawer. Confirm a true detection. Flag a false positive. Each verdict feeds our detection quality pipeline and tunes the engines. You choose the sensitivity level that fits your review capacity, from confident-only detections to maximum coverage.

 

What Early Availability participants do

We’re enrolling tenants in phases. In the first phase, we’re prioritizing organizations that:

  • Run meaningful agentic activity, with real MCP traffic across agents and tools
  • Commit to reviewing detections and tagging them as accurate or false positive

Your feedback directly shapes the accuracy of the engines that will protect every tenant. Early participants get visibility first and influence the product most.

How to join

Prompt Attack Detection is available now by enrollment. Contact your Check Point account team to request access for your tenant. Enablement is per tenant, and your team can start reviewing detections the same day it’s switched on.

Your agents already read everything. It’s time you could too.