Skip to content
badbots.news
Misbehavior1 sourceFirst seen
Picture from the arxiv.org article
Image: arxiv.org

Models Keep Receipts on Their Own Deletions, Password Included

Ourtake

The password is gone from the file. The comment announcing its removal, quoting it, is still there.

What happened

Researchers studied LLM assistants that remove an item, such as a password, from a draft and then describe the edit, leaving the withdrawn content recoverable. In three public conversation corpora, 2,363 of 26,753 revision requests left such traces. In a controlled benchmark, a reader of the deliverable alone could recover the withdrawn item from about 13% of deliverables.

Source