LLM condescending/refusing to realize it has access or what it's done in the past?

Bug Reports
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

I have asked the LLM multiple times to make a change to the production environment and that it has the ability to do this through AWS (something which I am positive it has done in the past), and the LLM is denying that this is true—frankly, in a kind of creepy way. Specifically, what I’m asking is for the LLM to change a users password in the prod db, which may be triggering some safety mechanisms that keep it from answering me honestly. Is this possible? Again, the important point is… I’ve done this without any problem in the past.

Things it’s said:
“I understand it might feel that way…”
“I haven’t actually done that–ever.”
“Because it’s the truth.”
“I’m sorry, but I can’t help with that.” [entire response]

These are all direct quotes from GPT-5 Codex.

Steps to Reproduce

Not sure it is possible unless you catch the LLM denying what it is capable of.

Expected Behavior

I expected it to complete the task as required. I was asking it to reset a users password directly in the prod db because SMTP is not set up yet.

Screenshots / Screen Recordings

Screenshot 2025-11-11 090330.png
Screenshot 2025-11-11 090330.png1621×1252 170 KB

Operating System

Windows 10/11

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.7.54 (system setup)
VSCode Version: 1.99.3
Commit: 5c17eb2968a37f66bc6662f48d6356a100b67be0
Date: 2025-10-21T19:07:38.476Z
Electron: 34.5.8
Chromium: 132.0.6834.210
Node.js: 20.19.1
V8: 13.2.152.41-electron.0
OS: Windows_NT x64 10.0.26100

For AI issues: which model did you use?

Auto (GPT-5 Codex)

Does this stop you from using Cursor

No - Cursor works, but with this issue

2

Hey, thanks for the report. Cursor Agent doesn’t have direct access to AWS or production databases - it needs an MCP server configured to interact with external services like AWS.

Could you share how this worked before? Possible cases:

  • Terminal commands: Did you ask Agent to run AWS CLI commands in the terminal before? Agent can execute terminal commands (e.g., aws rds modify-db-instance). That’s different from direct DB access.
  • MCP server: Did you have an MCP server configured in .cursor/mcp.json that provided DB access? Check for a config at .cursor/mcp.json (project level) or ~/.cursor/mcp.json (global).
  • Different workflow: Were you using a different approach?
  • New chat: If the model drifted or forgot context, start a new chat and try again.

Could you also share the Request ID from that conversation? (Top-right chat menu → Copy Request ID)