The AI Protocol Nobody's Watching Could Be Your Biggest Security Nightmare

Let's talk about MCP, the Model Context Protocol.

If you're deploying AI agents in your organization, and a lot of you are, or your leadership is asking why you aren't, this is the story you need to stop and actually read.

A security researcher named Syed Anas Mohiuddin published findings on a class of attack called Protocol Pivoting.

What that means in plain English is this.

Your AI agents are designed to trust each other.

Agent A gets an instruction, passes it to Agent B.

B passes it to C.

Clean and efficient.

except when someone injects malicious instructions into that chain because the guardrails that would stop a bad prompt at the LLM level, they don't apply between agents.

The agents just trust each other like a really enthusiastic intern who forwards every email without reading it.

He found these vulnerabilities across Google, JPMorgan Chase, Weevy8, Rapid7, and the French government's digital directorate.

Google's exposure scored an 8 out of 10 in severity.

That's not a paper cut.

That's a real wound.

Their MCP database toolbox wasn't validating IP addresses or applying redirect policies, which opened the door to server-side request forgery.

Rapid7's was only a 2.7 on the severity scale.

Sounds low, right? But it still allowed an attacker to plant malicious texts that agents would pass down the chain like a normal delegated task.

Rapid7 patched theirs last month.

Good for them.

The broader ecosystem? Still exposed.

Here's why this matters specifically to AEC firms and construction tech teams.

you are in the middle of a land rush right now.

Everyone's trying to stand up agenic workflows, AI systems that can read documents, make decisions, take actions, RFIs, submittals, cost tracking, scheduling, all of it.

That's genuinely exciting.

I've been building in this space and the potential is real.

Let's talk about MCP, the Model Context Protocol.

If you're deploying AI agents in your organization, and a lot of you are, or your leadership is asking why you aren't, this is the story you need to stop and actually read.

A security researcher named Syed Anas Mohiuddin published findings on a class of attack called Protocol Pivoting.

What that means in plain English is this.

Your AI agents are designed to trust each other.

Agent A gets an instruction, passes it to Agent B.

B passes it to C.

Clean and efficient.

except when someone injects malicious instructions into that chain because the guardrails that would stop a bad prompt at the LLM level, they don't apply between agents.

The agents just trust each other like a really enthusiastic intern who forwards every email without reading it.

He found these vulnerabilities across Google, JPMorgan Chase, Weevy8, Rapid7, and the French government's digital directorate.

Google's exposure scored an 8 out of 10 in severity.

That's not a paper cut.

That's a real wound.

Their MCP database toolbox wasn't validating IP addresses or applying redirect policies, which opened the door to server-side request forgery.

Rapid7's was only a 2.7 on the severity scale.

Sounds low, right? But it still allowed an attacker to plant malicious texts that agents would pass down the chain like a normal delegated task.

Rapid7 patched theirs last month.

Good for them.

The broader ecosystem? Still exposed.

Here's why this matters specifically to AEC firms and construction tech teams.

you are in the middle of a land rush right now.

Everyone's trying to stand up agenic workflows, AI systems that can read documents, make decisions, take actions, RFIs, submittals, cost tracking, scheduling, all of it.

That's genuinely exciting.

I've been building in this space and the potential is real.