Mackenzie Brown

AI agents are arriving in your clients’ environments faster than anyone can write the rules for them, and attackers are already using them. In 2026, operators pointed agents at a government and reportedly compromised 85 accounts in four days, with no zero-day

needed. As the channel starts writing standards for managed intelligence (for AI workloads), MSPs need a working understanding of how agentic attacks actually work, and why an MSP’s delegated admin access makes it the most valuable trust hub in the chain.

This session is a 101. We’ll start with the vocabulary that matters: the “lethal trifecta” of private data, untrusted content and an exfiltration path, and the OWASP Agentic Top 10 risks. In the Intel Room, you’ll apply what you learned, voting live as a real AI-driven campaign unfolds.

You’ll leave with a five-control starter baseline drawn from this year’s NSA, CISA and NIST agent guidance: an agent inventory, scoped and short-lived credentials, human sign-off on destructive actions, tool-call logging and a kill switch.

Learning objectives:

• Define the lethal trifecta and the core OWASP agentic risks in plain language

• Explain why MSP delegated access turns one AI agent compromise into many client compromises

• Recognize four early signals of agent abuse: agents launched with approval prompts bypassed, agents run by service accounts, agents reading stored credentials, and AI service traffic using a key that doesn’t belong to that host or user

• Apply a five-control agent baseline to a client environment