Attacking & Threat Modeling The Agentic Top Ten: ASI04 - Agentic Supply Chain Vulnerabilities
The “hopium” phase of AI security is officially over. We have to stop pretending that we can simply “patch” our way into a secure agentic ecosystem.
Anthropic’s latest admission—that agents are fundamentally unsecurable due to data poisoning and opaque supply chains—is a wake-up call for any European enterprise moving beyond the pilot phase. In my view, the risk isn’t just a single malicious actor; it’s the fact that agentic systems compose their capabilities at runtime. When you allow an agent to dynamically pull in third-party tools, MCP servers, or “personas” on the fly, you aren’t just deploying software—you are opening a live door into your infrastructure where the dependencies change every millisecond.
For the public sector especially, this means traditional perimeter security is obsolete for AI. We need to shift immediately toward Zero Trust architectures and rigorous runtime orchestration. If you can’t see or control the component at the moment of execution, you cannot trust the output. It’s time to move from “how do we make this agent smart?” to “how do we contain an agent that is inherently compromised?”
How are your teams currently validating the third-party tools and “skills” being plugged into your AI workflows?
#AISecurity #AgenticAI #CyberSecurity #PublicSector #ZeroTrust
https://disesdi.substack.com/p/attacking-and-threat-modeling-the-603