Embedded AI: Evolving Attack Surface And Ways To Defend Them

## Beyond the AI Buzzword

This WICCON talk distinguishes traditional AI, generative AI, and embedded AI—models that run at the edge without sending data to remote servers. That local processing embeds intelligence directly into workflows, often with little human validation of outputs.

## Growing Attack Surface

The speaker reviews market growth drivers and emerging risks: large interconnected datasets tied to sensitive information, reduced human oversight, and new integration points attackers can abuse. Edge AI expands the blast radius when models or their pipelines are compromised.

## Real-World Example: MCP

A highlighted case is the Model Context Protocol (MCP), rapidly adopted to connect AI systems to internal data. An RCE flaw (since patched) allowed arbitrary command execution when developers visited malicious sites, stemming from weak input validation and missing authorization checks—a pattern seen across many AI deployments.

## Defensive Guidance

Organizations should treat AI components like any other critical software: enforce access controls, validate inputs, monitor model behavior, and assess embedded deployments separately from cloud LLM use cases. Security must keep pace with how fast AI is being wired into production systems.