LEM: what data was the AI exposed to?
LLM Exposure Monitoring
An agent’s output tells us only part of what happened. This project studies how to record the data it reads, the APIs it uses and the information entering its context across different platforms.
A correct answer can still involve sensitive data.
A useful summary does not reveal which documents entered the context. To review that exposure, a monitor needs access records with clear event meanings and sources.
Open recording patterns
The paper proposes shared exposure-event semantics, including agent attribution and attestation sources, and examines the gap between actual exposure and what platforms make observable. Explore the four exposure surfaces below.
Read
Data read from files, databases and other platform resources.
Write
Data the agent creates or modifies on a platform: what changed, as well as the fact that a change occurred.
API
Requests and responses exchanged with external services.
Context
Information entering the model’s context and exchanged with its provider, including documents, conversation history and returned outputs.
Open Recording Patterns for LLM Exposure Monitoring: A Gap Analysis and Schema Proposal
Alex Chengyu Li · Preprint
The public experiment repository contains a dual-stream corroboration engine, synthetic-event experiments and replication data. The paper distinguishes simulation results from a GitHub webhook pilot.
Related project: AEGIS studies how recorded actions become independently verifiable evidence.AEGIS