AI Pre-flight: Mapping the Agentic AI Manipulation Defense Landscape and Proposing a Write-Time Detonation Architecture
Year: 2026
Type: preprint
Abstract: A morse code message drained a crypto wallet of approximately $175,000. Not a sophisticated exploit. Not a nation-state operation. A string of dots and dashes embedded in content an AI agent read, interpreted as an instruction, and acted on. The agent did exactly what it was designed to do. That was the problem. Agentic AI manipulation works the same way every time. Something the agent reads steers what it does. The damage is done, or in a particularly humorous case, a flan recipe. Input filtering, provenance tracking, and behavioral rules have all been proposed against this class of attack. They work to keep the agent from being manipulated. What it does after that goes unchecked. AI Pre-flight intercepts every action an agent attempts before it executes. The approach is borrowed from malware sandboxing: detonate first, evaluate what comes out. What gets evaluated is the blast radius. The effect, not the cause. We step away from trying to guess what malicious intent looks like, and focus on malicious results. Today it is morse code, tomorrow it is Klingon. By focusing on the end state, it does not matter what the attack vector was.
Author: Lozano, Mauricio
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162004346
Cites:
Cited by:
Related to: