liveBook · Manning

User input leakage

As explained in Building Reliable AI Systems

A user-input leak is the exposure of sensitive information supplied by a user to an AI system.

User-input leakage occurs when sensitive information entered into an AI system is exposed through weak session management, logging practices, or fine-tuning procedures. Users may share personal or financial identifiers, proprietary business data, and confidential communications believing their conversations are private. section 11.5.1

Cross-session exposure

An early ChatGPT caching bug exposed parts of some users’ chat histories to other users, illustrating how session handling can leak conversation data. section 11.5.1

How Building Reliable AI Systems covers it

How deep it goes
Names session management, logging, and fine-tuning as leak paths, but does not detail mitigations.
Examples
Cites an early ChatGPT caching bug that exposed parts of users’ chat histories across sessions.
What you can do after
Understand how sensitive user data can leak through session handling, logging, and fine-tuning.
How it connects
Connects to session management, logging, fine-tuning, and privacy failures.
Cover of Building Reliable AI Systems
liveBookmanning.com
Read Building Reliable AI Systems on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Building Reliable AI Systems. The section links point to the parts of the book it draws on.