ForgeNEX

Anthropic Extension in Cursor, 12-Hour Backdoor

An official Claude extension in Cursor was compromised and affected up to 5,000 users. We analyze the real risk for SMBs and what to do today.

4 min read

In this article

A compromised official channel: when trust is not enough

Last week, an incident came to light that should make any IT team reflect: an official Anthropic extension for the Claude model, distributed through the Cursor marketplace, was tampered with to include a backdoor. The malicious version, 2.1.313, was available for nearly twelve hours and, according to initial estimates, may have affected up to 5,000 users, presumably at the time they downloaded it or when their systems installed it automatically.

The key detail is that the package was published from Anthropic's official account on Cursor. It was not a clone or a similar name: it was a replacement of version 2.1.312 with 2.1.313 on the legitimate channel. Anthropic later released version 2.1.314, but neither the company nor Cursor has yet explained how the attackers gained control of the account. Nor is there any record that Anthropic notified its customers of the incident, according to sources consulted by Ars Technica.

Illustrative detail: Anthropic VS code extension on Cursor backdoored for nearly 12 hours

What the attackers were after

According to the CTO and co-founder of the security firm Phorion, Luke Roberts, the goal was to obtain cloud credentials. This is a common pattern in supply chain compromises: the victim is not attacked directly, but rather the tool they install with trust in order to reach their secrets. In this case, the vector was an AI-assisted development extension, a type of software that more and more teams install without going through a formal review process.

Independent researcher Kevin Beaumont and Phorion agree that version 2.1.313 was distributed through Anthropic's official account on Cursor. Automatic updates are enabled by default in Cursor, although users can configure this in the settings. That detail is important: many machines did not need any human action to receive the compromised package.

Why this matters to an SMB

At first glance, an incident in an AI development extension may seem like a problem for large technology companies. It is not. SMBs have adopted generative AI tools at a very high rate, and many have done so without a clear policy on which extensions, plugins, or assistants can be installed on work devices. The result is an attack surface that grows faster than the controls.

There are three factors that turn this case into a wake-up call for any organization:

  • The channel was official. There was no fake domain or phishing email. The extension came from the correct site, which invalidates the idea that "if it is in the marketplace, it is safe."
  • The update was automatic. Within hours, thousands of machines may have run unwanted code without anyone intervening.
  • The target was cloud credentials. A compromised development laptop can provide access to repositories, databases, billing services, or production environments.

This incident connects directly with what we already warned about when we discussed that SHA pinning is no longer enough: the software supply chain is an open front, and relying solely on the signature or origin of the package does not cover all scenarios.

What we recommend doing starting today

This is not about banning AI in development, but about introducing proportionate controls. These are the measures we suggest to our clients:

  1. Review installed extensions. Take an inventory of the VS Code, Cursor, and similar tool extensions on the team's devices. Identify which are official, which are not, and who authorized them.
  2. Disable automatic updates in critical tools. This can be configured in Cursor. Establish a review window before applying updates in sensitive environments.
  3. Limit and rotate cloud credentials. Apply the principle of least privilege to the access keys used by developers and rotate any that may have been exposed.
  4. Monitor anomalous access. Review your cloud provider logs for access from unusual locations or devices in recent weeks.
  5. Train the team. Make sure developers know that a marketplace extension may be compromised and that they should report strange behavior without fear of reprisals.

This case also reinforces the idea we defend in cybersecurity and AI: the investment that boosts your results: AI brings productivity, but it requires a proportionate layer of security. Without it, time savings become a risk that no one has budgeted for.

The lesson: verified trust, not blind trust

The incident of the Anthropic extension in Cursor leaves an uncomfortable but necessary idea: official channels also fail. For an SMB, the answer is not to stop using AI, but to assume that any external component—an extension, a plugin, a dependency—must be treated as third-party code that has to be governed. That implies inventory, version control, credential management, and response capability. Twelve hours of exposure may seem short, but with automatic updates they are enough for an attacker to get very far.

Source: Ars Technica. Analysis and adaptation: ForgeNEX.

Keep reading

Call us WhatsApp