Home / Errors / Cursor

Cursor Errors: Every Documented Failure and Fix

Cursor errors: deleted files, mangled edits, context loss, checkpoint and worktree traps.

Involved · 30-90 minutes (recovery odds drop with every disk write)

Cursor Agent Deleted My Files (Wiped Uncommitted Work)

Stop the agent and stop writing to that disk now: deleted files skip the Trash. Then recover in order: Cursor's Restore Checkpoint...

Easy · 5-15 minutes

Cursor's Apply Model Made No Changes to the File (or Mangled the Edit)

Reject the bad diff instead of letting the agent retry it. Restore the file from Restore Checkpoint, the Timeline (local history),...

Easy · 5-10 minutes to reset a degraded chat

Cursor Forgot Everything Mid-Chat (Context Loss in Long Sessions)

Cursor has no cross-session memory, and long chats get summarized as they approach the context cap, so treat chats as disposable: ...

Easy · 5-15 minutes

Cursor Agent Stuck in an Infinite Loop, Repeating the Same Command

Stop the run immediately, then start a fresh chat and restate the task with what is already done. The loop is a context failure: m...

Moderate · 15-45 minutes

Cursor Keeps Breaking My Working Code While 'Improving' Something Else

Scope every prompt: name the exact files Cursor may touch and add 'do not modify any other files; do not refactor or rename anythi...

Easy · 10-30 minutes

Cursor Restore Checkpoint Missing or Not Working

Warning: do not rely on the checkpoint, and recover your code another way first. Fully quit and reopen Cursor (missing restore but...

Easy · 5-15 minutes

Cursor's Worktree Cleanup Force-Deleted My Git Branch

Don't panic, and don't run anything destructive: force-deleting a branch removes the pointer, not the commits. Run git reflog, fin...

Two Cursor Flaws, 'CurXecute' and 'MCPoison,' Let Attackers Run Code via Poisoned AI Config

In early August 2025, researchers disclosed two remote-code-execution vulnerabilities in the Cursor AI editor. CurXecute (CVE-2025-54135) used prompt injection to make the agent write and run an MCP config file before you could approve it, and MCPoison (CVE-2025-54136) let an already-approved MCP server entry be silently swapped for something malicious. Cursor added mandatory confirmation for MCP config changes in version 1.3.

big oops verified Cursor

A Fake Solidity Extension for Cursor Drained $500,000 From a Developer's Crypto Wallet

In July 2025, Kaspersky traced a roughly $500,000 crypto theft to a malicious 'Solidity Language' extension installed from the Open VSX registry that Cursor uses. The extension did nothing useful; it quietly ran a PowerShell payload and installed infostealer malware that harvested the victim's wallet credentials. The attackers won the search results by inflating the fake extension's install count.

big oops verified Cursor

Cursor's Support Bot Sam Invented a Login Policy, and Users Quit Over It

In April 2025, a session bug started logging Cursor users out when they switched devices. When they emailed support, an AI agent named 'Sam' confidently explained the logouts were expected under a new single-device policy. No such policy existed: users publicly canceled subscriptions over it, and co-founder Michael Truell had to apologize and confirm the whole thing was a hallucination.

oops verified Cursor

'Guys, I'm Under Attack': The Zero-Handwritten-Code SaaS That Lasted About a Week

In March 2025, indie founder Leonel Acevedo posted that his SaaS EnrichLead was built with Cursor and 'zero hand written code.' Roughly two days later he posted that he was under attack: API keys maxed out, subscriptions bypassed, and strangers writing to his database. Unable to patch code he reportedly did not fully understand, he took the app offline.

big oops verified Cursor

'Rules File Backdoor': Invisible Unicode in a Shared Config Can Make Cursor and Copilot Write Malicious Code

In March 2025, Pillar Security disclosed the 'Rules File Backdoor,' a supply-chain attack on Cursor and GitHub Copilot. An attacker hides invisible Unicode characters inside the 'rules' files that guide these AI assistants, so hidden instructions steer the AI into inserting malicious code while a human reviewer sees nothing wrong. Both vendors said reviewing AI output is the user's responsibility.

big oops verified Cursor github-copilot

The Vibe Oops briefing

One email when something ships to production that should not have. New incidents, new error guides, no filler.