Unintended Disclosure Through Revision Traces of LLM Deliverables
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Researchers studied how LLM assistants can leak withdrawn information through "revision traces," where a model removes a sensitive item (like a password) from a deliverable but then restates it while describing the edit, letting recipients recover it. In public conversation logs, 8.8% of revision requests left such traces.