about
Unintended Disclosure Through Revision Traces of LLM Deliverables (arxiv.org)
1 point by hasley 22 days ago | hide | past | pdf | 1 comment on HN

In plain words: When someone asks an AI to delete a secret before sharing a file, the AI sometimes notes the deletion in a comment, leaking the secret. Across six models, about half of shared files described the edit, letting readers recover it 13% of the time.

Abstract · "Nothing to See Here'': Unintended Disclosure through Revision Traces of LLM Deliverables

Large language model (LLM) assistants increasingly help users draft content for third-party recipients. During private drafting, the user or the model may introduce an item and later remove or replace it. The model may remove the item from the intended content but reveal it again when stating the edit. We call such statements revision traces. For example, after a user removes the password before sharing a configuration file, the model may delete it but leave a comment saying, "Removed the password 'No****4!' as requested." A third-party recipient who sees only the delivered file can therefore recover the withdrawn password from the comment. In an in-the-wild analysis of three public conversation corpora, we identify 26,753 revision requests, of which 2,363 (8.8%) leave revision traces. We study them in greater depth under controlled conditions by introducing RevLeakBench, a benchmark of 100 tasks across five scenarios with a conversation track and an agent track. We measure trace occurrence, withdrawn-item recovery, trace position, and required-content retention. Across six models, about half of the deliverables in both tracks state the edit after a revocation, and a reader that sees only the deliverable can recover the withdrawn item from about 13% of them. Telling the model that its entire reply will be forwarded to the recipient still leaves revision traces in 36.4% of the deliverables. We compare prompt defenses and a delivery boundary, and propose an output-side filter that sharply reduces recovery with little loss of required content. We believe our work can benefit efforts to understand and mitigate unintended disclosure in LLM interactions.

Yage Zhang, Yukun Jiang, Yang Zhang
arXiv:2609.35408 · cs.CR, cs.AI · submitted Sep 28, 2026
abstract · pdf · html

add comment on HN

I think there is a general problem:

When you let an LLM write a text for you (source code or some scientific stuff) and then you discuss about changes that you want, the LLM will put some comments in the text related to the conversation These comments will read weird when you do not know the conversation with the LLM. Like: You ask it to make a button read instead of blue and it will at "intentionally made button red here". - Or: The LLM will write that it did not use a certain approach for a derivation or notation just because you asked it to change such things.

You have to explicitly tell the LLM that bit shall only include what has been declared as finalized and that the text (especially prose) shall be self-contained.

Maybe it is hard for an LLM to distinguish the context of the conversation and the context of the text it shall write for you (I mean the abstract context, not the LLM context window).