AckuityRequest access WHITEPAPER · 8 PAGES
A taxonomy of threats to AI agents.
A framework for examining the new attack surface created when AI systems interact with data, use privileges and take actions.
THE RESEARCH
Four classes. Twenty-one threat vectors.
The paper organizes threats to agents in production into four broad classes: data interaction, privilege misuse, agent manipulation and overreach. It provides a foundation for identifying risks beyond the contents of an individual prompt.
Use it to frame architecture reviews, discuss control coverage and examine the points where an agent’s proposed action needs further context.
Download whitepaper · PDFTake the research into your architecture.
Explore the broader threat catalog or learn how independent verification fits at the execution boundary.
NEXT STEP
Turn a threat model into a practical review.
Explore what these risks mean for your agents with read-only discovery.