The Meta AI support issue is a useful warning for anyone building AI agents into support workflows.

The problem is not AI support itself.

The problem starts when AI can trigger sensitive actions like account recovery, access changes, refunds, or internal escalations without strong enforcement outside the model.

At that point, the AI agent is no longer just a chatbot.

It is part of the security boundary.

I think teams need to start asking harder questions before shipping these systems, like:

• Can the AI perform privileged actions?

• Can it be socially engineered?

• Are risky prompts scanned before tool execution?

• Are responses and actions logged?

• Is there human approval for high-impact workflows?

This is also why tools like LangProtect matter in the AI workflow layer. Prompts, responses, file uploads, and tool-triggered actions need real-time scanning and policy enforcement, not just better system prompts.

Are most teams actually threat modeling AI support agents today, or are we still treating them like normal support bots?