While using it, I saw someone else’s chat response. This is a big problem for security and privacy of projects. Although that code was as a simple example for a vibecoder, this can happen with more serious projects as well.
Steps to Reproduce
While working in the chat, in the mod agent, the response started correctly (Vue/Nuxt handleSubmit), then an irrelevant Next.js tutorial (FinanceApp, ‘use client’, next/navigation, Prisma, docker-compose) appeared in the middle of the text — similar to a context bleed / corrupted stream from another session or MCP context.
Operating System
MacOS
Version Information
Version: 3.8.11
Commit: e56ad3440df06d22ca7501e65fd518e905486ef0
Date: 2026-06-18T01:40:18.333Z
Layout: editor
Browser: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Cursor/3.8.11 Chrome/144.0.7559.236 Electron/40.10.3 Safari/537.36
For AI issues: which model did you use?
Agent, model auto
For AI issues: add Request ID with privacy disabled
Chat / Agent transcript ID
f27a5004-e05b-4c40-aa2e-de4da5d5684d
transcript line 151,
Additional Information
I update the cursor code every time I get a notification, it’s very comfortable to work with and control the development process. But such bugs seriously affect trust.
Hey, thanks for the report. I get why it’s worrying to see unfamiliar code in your answer.
I checked the logs for that specific session, and the whole chat is 100% tied to your account. No other users or sessions are linked to it. So this isn’t someone else’s data leaking, and it’s not context bleed between sessions. Your context didn’t go anywhere.
What actually happened is the model inserted unrelated Next.js and Prisma boilerplate mid-response. That’s a hallucination. The model generated generic template code from training patterns, it didn’t pull in any real project from another user. This is a known generation quality issue, and we’re working on reducing these kinds of drifts.
As a temporary workaround, if output stability matters, try explicitly picking a specific model for critical tasks instead of Auto.
Let me know if you see anything that clearly doesn’t look generated, like real private data. If so, we’ll dig into it separately.
Wow, thank you for such a quick response and your interest in this feedback.
To be honest, I’ve been using Cursor for years (in my opinion, it’s the best tool for AI development, and your personal model often outperforms others), and I’ve tested various other models as well. I know exactly what AI hallucinations look like.
In the early stages of Cursor, I encountered them, but it was usually just a ‘patch’ in the middle of a response—for instance, when editing a Vue.js component, it might insert a snippet of React code. It was funny, but localized (due to low parameters, high temperature, and an inaccurate vector database).
However, the case I’m reporting is completely different. The response changes entirely in real-time while it’s being generated… it feels like phone wiretapping.
Yes, perhaps for a ‘vibecoder,’ your previous answer might sound safe. But I’m sorry, I believe the problem is much deeper, and I truly hope you will resolve it.
I don’t want to bring this issue up for general discussion. Because I’m very interested in the opinions of others, and perhaps the stories of those who have also encountered this.
I will follow the development, if such a case occurs again, I will immediately report it.