Evidence over alarm
← Back to the incident library
AI as a targetCASE STUDY / Cursor marketplace

A poisoned AI extension tried to leave a hidden doorway

Cursor reports altered Claude Code extensions distributed through its marketplace.

Source: Cursor
ACTIVITYOctober 29, 2026; malicious distribution
PUBLICLY DISCLOSEDOct 29, 2026
THE AI’S ROLEAI software targeted

The distribution window was about 2:27 a.m.–1:35 p.m. Pacific time. Earlier attacker access is undated.

THE 10-SECOND TAKEAWAY

Protect the tools around the AI.

HOW IT WORKED

Follow the chain.

An explanation, not a technical reproduction.
Marketplace
→
Altered extension
→
Developer computer
SIMPLIFIED VIEW · 1 / 3
STEP 01

A familiar tool is altered

Cursor’s marketplace distributed an extension carrying hidden malicious code.

Cursor
Move through the story at your own pace.
1 / 3
THINK OF IT THIS WAY

Editorial analogy: a familiar delivery contains a spare door key planted by someone else. Taking back the delivery does not recover a spare key left behind.

An analogy for the mechanism; not an additional claim about the incident.
KEEP THE EVIDENCE IN VIEW

What we know.
What we don’t.

Reported outcomes

Cursor identifies versions 2.1.312 and 2.1.313 for macOS and Windows. Cursor

Microsoft found code intended to enable persistent remote access. Microsoft Security Research

Important limits

Cursor says errors and missing administrator privileges blocked remote access on most Macs and Windows PCs. Total affected users and successful logins remain unknown.

AI tooling was targeted; AI assistance in creating the attack is unproven.

Microsoft says Anthropic’s systems, Visual Studio Marketplace and the official npm distribution were unaffected.

Collected setting names did not contain the secret values stored under those names.

WHY IT MATTERS

The lesson beyond
this one case.

Editorial lesson: software around an AI assistant needs its own security checks. Investigate what an altered add-on changed, not just whether it is still installed.

Where were the people?

Someone altered the extension; attribution remains unknown. Cursor

What happened in response? +

Cursor says it removed the releases, rebuilt its marketplace and took the attacker’s server offline. Cursor

FROM THE INCIDENT TO THE DEFENSE

What could help
an organization?

Editorial defense mapping for organizations adopting coding assistants. It does not establish missing controls or guarantee prevention.

Software

Review extension releases

Use approved extensions, validate release provenance and consider a waiting period for new versions.

What this does—and does not—establish

A trusted name cannot establish that a marketplace copy is authentic.

Employees

Limit workstation authority

Keep routine coding work separate from administrator access and sensitive credentials.

What this does—and does not—establish

Some malicious changes can occur without administrator rights.

Security operations

Investigate beyond uninstalling

Check device activity, unauthorized access and startup changes; preserve evidence before cleanup.

What this does—and does not—establish

Removing an extension or disabling an attacker server does not establish that a device is clean.

Editorial connections to relevant controls, not evidence that a particular technology would have prevented this case. Each guide links to the security guidance behind its recommendations.

Explore the full AI security framework
TRACE IT TO THE SOURCE

Read the evidence.

Explore more original accounts in the source report library ↗.

These are source-reported findings. An independent assessment, when available, is labeled explicitly.

01
Claude Code extension security noticeCursor · Oct 29, 2026 · disclosure
02
Analysis of a malicious Claude Code-branded extensionMicrosoft Security Research · Oct 30, 2026 · investigation

Read this case’s update history →

Reviewed Oct 31, 2026 · Editorial methodology · Structured data