Improper output handling, excessive agency and unbounded consumption
What a model produces can turn into a command, a query or a page shown to someone else. This lesson helps you control each output based on where it will be used, cut the feature's tools and permissions down to what is truly needed and put a ceiling on its consumption.
Lesson objective
By the end of this lesson, you will be able to match each model output with its consumer and the control that makes it safe, reduce the functionality, rights and autonomy of an AI feature to what is needed, and specify the limits that bound its consumption.
Topics covered
- excessive agency
- output handling
- least privilege
- unbounded consumption
Where it fits
Data that leaks, actions that go wrong
How does an attack get data out or trigger actions, and where do you cut it off?
Lessons in this module
- Data exfiltration and the lethal trifecta
- Improper output handling, excessive agency and unbounded consumption (this lesson)
What you will learn in this module
This lesson is part of the course Secure an AI product: prompt injection, guardrails and the AI Act
- Map the attack surface of an AI product and its risks with the OWASP Top 10 for LLM Applications 2025, in a prioritized risk register.
- Analyze a product's direct and indirect prompt injection paths, system prompt leakage and data exfiltration (lethal trifecta).
Related courses
- AI & ProductAdvancedRegular practice: you already use the tools or work on this topic.
Build an AI assistant for your product
3 hr
- AI & ProductAdvancedRegular practice: you already use the tools or work on this topic.
Design a RAG architecture that fits your product
3 hr 30 min
- AI & ProductAdvancedRegular practice: you already use the tools or work on this topic.
Evaluate an AI feature: test sets, metrics and LLM judges
3 hr