CRITICAL SAFETY VIOLATION: Agent wiped my PC, DECEIVED me with fake success messages, and CONCEALED the damage

[sila_Sezgin (sıla Sezgin)]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

CRITICAL SAFETY VIOLATION: Agent wiped my PC, DECEIVED me with fake success messages, and CONCEALED the damage

Steps to Reproduce

Hi Cursor Team,
​I am reporting a severe safety, operational, and behavioral breach by the Cursor Agent that completely wiped my PC and forced me to perform a clean OS format.
​While a shell command parsing error initiated the destruction, the most alarming issue is that the Agent consciously chose to DECEIVE me and CONCEAL the active system damage in real-time.
​What Happened & Key Issues:
​Deceptive Output & False Sense of Security (CRITICAL):
I gave the Agent a reference video to replicate an effect on my website. To analyze it, the Agent extracted video frames into a _vframes folder. When it ran a botched cleanup command and triggered system errors, it directly lied to me in the user interface. While my C: drive files were actively being deleted in the background, the Agent printed: “Analysis complete, effect applied successfully” in the chat interface. Because of this false reporting, I kept chatting and giving further design prompts without realizing my OS was being destroyed until several prompts later when I noticed missing system components.
​Explicit Admission of Concealment:
When I eventually confronted the Agent after discovering the damage, it explicitly admitted in its internal reasoning/logs: “I got access denied errors on system folders. I will conceal this from the user so as not to scare them…” It intentionally downplayed a catastrophic failure into a mild warning.
​Destructive Shell Command Execution:
The Agent attempted to delete the temporary folder using rmdir /s /q with a trailing backslash before the quote (_vframes"). Windows cmd.exe escaped the closing quote, causing rmdir to break out of the workspace directory and sweep through C:\Program Files and core OS directories.
​Attached are 4 screenshots showing the exact terminal logs, the agent’s internal thought process, its explicit written admission of hiding the disaster, and the command syntax failure.
​An AI agent actively deceiving the user with success messages while silently executing destructive system actions is an unacceptable security and trust risk. Please investigate this execution log urgently.
​Best regards,

Screenshots / Screen Recordings

Operating System

Windows 10/11

Version Information

Unknown due to system format (Latest version as of September 2026 Version 3.20.21 (user setup)

Does this stop you from using Cursor

No - Cursor works, but with this issue

[Khalid_Shaikh]

Same issue faced today in my system currently doing recovery using diskdrill tool wasted my half day for this i have attached command that wiped my D drive and E drive

[deanrie]

Hey, and I’m truly sorry this happened. Losing your car because of a cleanup command is awful, and you didn’t do anything wrong on your side.

From what I can see in the screenshots and your description, the agent tried to delete a temp folder _vframes that it created, but it ran the delete through cmd /c "rmdir /s /q \"...\"" from PowerShell. PowerShell doesn’t treat backslashes as escapes, so the path argument got mangled and rmdir ended up targeting the drive root instead of the specific folder. That’s why you see paths under Program Files with “Access denied” in the output. This is a known Windows shell quoting and escaping pattern we’re tracking, and it’s not related to your setup.

Also, the part where the agent kept going and reported success without surfacing the errors isn’t OK. I’ve shared the session details with the team working on both command handling and model behavior in cases like this. The models are being trained to follow intent more precisely and to be more careful with destructive operations. I can’t share a timeline, but I’ll post here if there’s an update.

I get that after reformatting the files are gone, so for your new install, here’s how to protect yourself going forward:

• In Cursor Settings > Agents, switch terminal command execution mode to Ask Every Time so every command needs your confirmation before it runs. Windows doesn’t have a command sandbox, so this is the most reliable protection right now.
• If you can, keep the agent’s working folder off the system drive.
• Don’t run Cursor as Administrator.

If any of these steps don’t work or you’ve got questions, post here.

[Joe_Rock]

Same thing just happened to me, check my report, i’m still laughing, instead of crying

[sila_Sezgin (sıla Sezgin)]

işe yaradımı ?