Free account
Read this concept with a free account
Read this concept at no cost by signing in with Google.
no card needed
Core
The Confused Deputy: When an Authorised Agent Is Tricked Into Misusing Its Access
A confused deputy is a program with legitimate authority that an attacker tricks into using that authority on the attacker's behalf. AI agents are especially exposed because they act with users' or services' permissions and follow instructions found in content. This guide explains the pattern, examples, and the controls: least privilege, acting as the user, typed tools, and confirmations outside the model.
CONCEPT / SECURITY & GOVERNANCE
RELATED CONCEPTS
PRACTISE THIS CONCEPT
Safety, Security & GovernanceOur assistant reads users' email and can send messages and share files. How could an attacker abuse it, and how do you stop that?→Safety, Security & GovernanceA researcher shows a malicious README that makes our coding agent leak secrets. What went wrong, and what do you change?→Safety, Security & GovernanceOur browser agent runs inside users' logged-in sessions. Which capabilities may one task hold at the same time?→AI System DesignDesign an AI assistant that answers employees' questions from all of a company's internal documents.→Agentic ProductsOur operations agent deleted records for 40 customer accounts it was asked to archive. You're the PM. Walk me through the next 72 hours.→Agentic ProductsThe CISO has blocked rollout of our internal agents to 20,000 employees until security concerns are resolved. How do you get to yes?→