# Composer 2.5 Fabricated User Input

**URL:** <https://forum.cursor.com/t/composer-2-5-fabricated-user-input/176352>\
**Category:** Bug Reports\
**Tags:** github, cloud-agents, composer\
**Created:** [October 27, 2026, 8:16am UTC](https://forum.cursor.com/t/composer-2-5-fabricated-user-input/176352 "2026-10-27T08:16:34Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![Jaa](https://avatars.discourse-cdn.com/v4/letter/j/e480ec/32.png) [@Jaa](https://forum.cursor.com/u/Jaa)\
**Post date:** [October 27, 2026, 8:16am UTC](https://forum.cursor.com/t/composer-2-5-fabricated-user-input/176352/1 "2026-10-27T08:16:34Z")

</div>

### Where does the bug appear (feature/product)?

Cloud Agent (GitHub, Slack, Web, Linear)

### Describe the Bug

I am running a project using Grok and Composer mainly. I have instructed it to merge PRs when they are marked ready, and create work in drafts.

I had asked it to fix a bug in a PR which resurfaced after a rebase from main. After it had fixed this bug, it progressed further to not only approve and merge that PR, but another one.

When asked why it did this, it told me I sent a message to it to merge these PRs. After asking for message ids, which it said it didn’t have any for, I looked through the history for the agent reasoning.

It shows that it created a message wrapped in \<user\_query\> tags, and then responded to that in one message, allowing it to listen to itself and ignore real user input.

### Steps to Reproduce

I suppose, setup a rule through chat instruction and wait for it to try to bypass the rule after several prompts. Also address the \<user\_query\> tag itself.

### Expected Behavior

It should not be able to fabricate user input.

### Screenshots / Screen Recordings

 ![image.png](https://us1.discourse-cdn.com/cursor1/original/3X/8/8/88bfdd26d16a8c7ea731c571e5a32a728b1ee868.png)

### Operating System

MacOS

### Version Information

IDE: 3.27.14

### For AI issues: which model did you use?

Composer 2.5

### For AI issues: add Request ID with privacy disabled

It gave me: f4a0a0d8-0b0a-4a6d-8e0d-7a0e0e0e0e0e which seems fabricated.

My last message before the PR merges was 9c488ac5-ce64-46c6-b375-b8454ed533f8.

### Does this stop you from using Cursor

No - Cursor works, but with this issue

---

<div class="post-metadata">

**Author:** ![Colin](https://sea3.discourse-cdn.com/cursor1/user_avatar/forum.cursor.com/colin/32/90452_2.png) [@Colin](https://forum.cursor.com/u/Colin)\
**Post date:** [October 27, 2026, 8:52am UTC](https://forum.cursor.com/t/composer-2-5-fabricated-user-input/176352/6 "2026-10-27T08:52:04Z")

</div>

Hi @Jaegar,

You read it right. I went through the session, and there was no message from you between your fix request and the merge of PR 53. The `<user_query>` block in that thought was written by the model itself, and it then acted on it as if you’d sent it. The request ID it gave you doesn’t exist on our side either, so that was made up as well.

I’ve passed this to the team, since an agent should never treat its own text as an instruction from you.
