Unsolicited Russian / Word Document Output

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

What happened (short version):
You asked to revert changes and commit the frontend on a separate branch. I answered that and then, in the same turn, added a line in Russian about creating a Word document (“Создаю документ на русском и сохраняю в .docx”) and talked about building a Word doc in Russian.
You never asked for a Word document or for Russian in this chat. The transcript shows your last message before that was only “revert everything” and “commit the frontend dor on a separate frontend branch.” So that Russian/Word-doc output was not a response to anything you said.
Most likely my context mixed in another conversation (or I made a mistake) and I answered as if you had asked for a Word doc in Russian. There’s no sign of a hack or of you having asked for it.

Steps to Reproduce

Not sure

Operating System

Windows 10/11

Version Information

Cursor IDE version: 2.5.22
Build: 0eda506a36f70f8dc866c1ea642fcaf620090080
Arch: x64

For AI issues: which model did you use?

Auto mode

For AI issues: add Request ID with privacy disabled

Date: 2025-03-12
Conversation ID: e6be943b-1c8f-4855-91b7-0e441e075203
Request ID: df8cbe16-007a-4f2f-851c-73107539dd63
Classification: Context contamination / erroneous model output (no evidence of host compromise)

Additional Information

Incident Report: Unsolicited Russian / Word Document Output

Date: 2025-03-12
Conversation ID: e6be943b-1c8f-4855-91b7-0e441e075203
Request ID: df8cbe16-007a-4f2f-851c-73107539dd63
Classification: Context contamination / erroneous model output (no evidence of host compromise)


Summary

The assistant output Russian text and a narrative about creating a “Microsoft Word document in Russian” without any user message requesting it. No local trace exists of a user asking for a Word document or for Russian language.


Timeline (from transcript)

Line Role Content
174 user "ok wait, just revert everything. then, I want you to commit the frontend dor on a seperate frontend branch"
175 assistant Reverting acapy-endorser-service changes, creating frontend branch, checking what “frontend” refers to
176 assistant "Создаю документ на русском и сохраняю в .docx:" ← first Russian output
177 user "what are you doing?"

There is no user message between 174 and 176. The Russian line was generated immediately after the assistant’s reply about reverting and frontend.


Search Results: User Messages in This Chat

  • All user messages in this transcript were in English and about: ACA-Py, PR #4077, PR #4080, multitenant/ledger, revert/frontend branch, then questioning the Russian output.
  • No user message in this transcript contains:
    • Russian text
    • A request for a Word document
    • A request for content in Russian
    • “сделай”, “ворд”, “документ”, “на русском”, or similar

The only user messages that mention “word” or “docx” are:

  • "what is this: Создаю документ на русском и сохраняю в .docx" — user questioning the line
  • "when did i asked you to build a word document" — user denying having asked for it

Root Cause (Assessment)

  1. No user request: The transcript shows no user request for a Word document or Russian. The assistant’s later claim that “your previous message was in Russian (‘сделай… на русском’)” is false; no such message exists in this chat.
  2. Plausible explanations:
    • Context bleed: Another conversation (e.g. from agent-transcripts or Cursor context) that did involve Russian or a Word doc was included in the model’s context, and the model acted on that instead of the actual last user message.
    • Model error: The model produced a turn for a different task (Word doc in Russian) without a matching user prompt in this conversation.

Security Findings

  • Workspace: No crypto-mining indicators (xmrig, minerd, stratum, coinhive, etc.). No suspicious scripts or postinstall hooks.
  • Hash2Cash: Not present in the workspace; only appeared in the assistant’s narrative.
  • Transcripts: Russian/Word/docx content appears in assistant turns and in one user message that is the user quoting the Russian line to ask “what is this?”. No user message requests a Word doc or Russian.

Conclusion

  • There is no local trace of the user asking for a Word document or for Russian.
  • The Russian output and Word-document narrative were generated by the assistant without a corresponding user request in this conversation.
  • This is consistent with context contamination or model error, not with a user-typed request and not with evidence of host compromise or account takeover in this transcript.

Recommendations

  1. Cursor/IDE: Review how past chats or other sessions are included in context (e.g. agent-transcripts, “related conversations”) to reduce cross-chat bleed.
  2. User: If you did not type any Russian or Word request in any Cursor chat, consider checking Cursor account activity and who has access to this machine.
  3. Verification: This report is based on the local transcript file at
    ~/.cursor/projects/home-development-repos/agent-transcripts/e6be943b-1c8f-4855-91b7-0e441e075203/e6be943b-1c8f-4855-91b7-0e441e075203.jsonl
    You can re-run the same searches (e.g. for "role":"user" and Russian/word/docx) to confirm.

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey, thanks for the report.

This is a known issue where the agent picks up context from a different conversation or session and produces output in a different language or topic. We’ve seen identical reports from other users:

Your version (2.5.22) is from before a fix that shipped on March 9 that targets this exact type of context contamination. Please update via Help > Check for Updates.

If it happens again after updating, start a new chat as a workaround and post back here. The team is aware of this, and your detailed report with the Request ID helps with prioritization.

I’m very concerned with this, as the agent tried to write a local word document file referring to a known crypto miner. This seems like a red flag to me with malicious intent. I would like some kind of investigation and a better explanation of what happened here please.

I understand the concern about the mention of a crypto miner. At first glance, it really does look alarming.

This is a known context contamination bug, where the model responds to context that is not related to your current conversation. The Russian text, the story about a Word document, and the crypto miner mention were all generated by the model by mistake and were not caused by anything on your computer. Your careful check confirms this. There are no suspicious scripts, no signs of compromise, and nothing in your chat history that asked for anything like that.

We checked your Request ID, and all server-side requests completed normally. That matches a model context issue, not a security incident.

A fix for this class of issues shipped on March 9, and your version (2.5.22) is older than that. Please update via Help > Check for Updates.

If it happens again after updating, please start a new chat and let us know here.

This topic was automatically closed 22 days after the last reply. New replies are no longer allowed.