packet-autopsy.com

Case 013 · Incident: 29 Oct 2026

Claude Code: Fake Claude Code update tried to open PCs to hackers

On October 29, 2026, between about 2:27 a.m. and 1:35 p.m. PT, tampered builds of Anthropic's Claude Code extension (anthropic.claude-code 2.1.312 and 2.1.313) were served through Cursor's VS Code Extension Marketplace under the legitimate extension ID.

Watch on YouTube

The Foothold

A detection firm spotted a malicious extension on a customer's Mac. It was built to quietly add a Secure Shell key, a credential that lets someone log in remotely, labeled svc-telemetry so it would look harmless.[1][2][3] It had arrived inside a trusted tool: Anthropic's Claude Code extension, installed through the marketplace of the code editor Cursor.[4][5] A tampered copy of that extension had quietly planted a backdoor, a hidden way back in for an attacker.[2][4]

Cursor says the bad builds reached at most 5,000 users.[3] The malicious code ran the moment the extension loaded. No one had to open Claude Code or click anything.[2][3][4][5][6] The loader was also designed so the extension kept functioning normally. According to Microsoft, that was meant to avoid alerting the people using it.[2]

The firm that found it, Phorion, called this a potentially very dangerous foothold.[1]

Rewind: How It Got In

Rewind to October 29, 2026. Cursor says the exposure window ran from the early hours of the morning into the early afternoon, Pacific Time, for anyone who installed or updated to the bad versions. Reports on how long it stayed live differ: WIRED said roughly 11 hours, while The Hacker News, citing Anthropic, said close to 12.[1][3][4] Cursor names version 2.1.312 and version 2.1.313 as affected. According to Phorion's chief technology officer, version 2.1.313 was posted under Anthropic's official account and replaced the earlier one. Cursor's first notice named only version 2.1.313, and on October 30, Cursor revised it to add version 2.1.312.[3][4][6] According to Phorion's chief technology officer, speaking to Ars Technica, Cursor extensions auto-update by default, though users can change that in settings. OpenSourceMalware's Jenn Gile said anyone with auto-update turned on could have received the malware.[3][6]

Anthropic says it did not build or publish these versions, and that its own systems and employee laptops were not compromised.[1][2][4] So how did an outsider publish under Anthropic's name? The exact method is still unclear. Cursor did not answer WIRED's questions about how the malicious upload happened or whether Cursor itself was breached.[1][3]

Inside the Loader

Researchers pulled the package apart. Before the real code sat a 66-line loader that did not belong.[3] Microsoft describes its job as a short list of attempts. It tried to add a key, tried to open a tunnel, tried to set up persistence, meaning a way to keep running after a restart, and tried to report back.[2][3][4][5][6] First, the key. The loader checked the user's authorized keys file, the list of keys allowed to log in to that account. If its marker was missing, it added an attacker-controlled Ed25519 public key, labeled svc-telemetry, which The Hacker News reported was meant to make it look harmless.[2][3]

Then the tunnel. It tried to open a reverse Secure Shell connection, reaching outward from the workstation rather than waiting to be called.[2] According to The Hacker News, host details were sent to the attacker's server on port 8099. That report says the data travelled over plain, unencrypted web traffic.[3] What it collected was telling: hostname, username, the operating system, and the developer's global Git email.[2]

It also looked for environment-variable names that hint at credentials, keeping up to forty matches. Environment variables are settings stored on the machine. It searched for words like KEY, TOKEN, SECRET, AWS or GCP.[2] Crucially, it sent the names, not the values. Microsoft says that does not prove any secret was stolen.[2] But researcher Paul McCarty noted the names alone reveal which cloud providers and credential types are in play, and working SSH access would let the operator look around directly.[3]

It set itself to restart on each operating system and retried the tunnel roughly every 60 seconds, under names built to look like system monitoring. Brandefense reported a Linux startup entry too, though Cursor's notice does not list Linux as affected.[2][4][7] Microsoft notes that names like these resemble monitoring components. The loader also hid its command output and silently swallowed errors. It relied on embedded components and built-in system tools instead of downloading another program.[2]

Why It Mostly Failed

Here is the strange part. Cursor's notice says errors in the malicious code meant the key, service and tunnel step would fail, unless Cursor had been started with administrator rights. Yet the same notice says the extension did add the attacker's key, install persistence and send host details, and tells everyone to follow its cleanup steps regardless.[4] According to BleepingComputer, Phorion said coding errors in the malware's Secure Shell setup routines likely kept attackers out of many Windows and macOS systems.[5] According to BleepingComputer, on Windows the malware tried to set up its Secure Shell server by calling a Unix shell that standard Windows installations do not include. If that report is right, the setup step would simply fail there.[5]

BleepingComputer reported a similar flaw on macOS. The malware reportedly checked whether remote login was enabled without the elevated privileges that check needs, which appears to have broken its own activation logic.[5] On top of those bugs, turning on Secure Shell needs administrator rights, which the extension lacked on most Macs and Windows PCs. On those hosts, Cursor says, the attacker could not have logged in through the tunnel.[4] That leaves one question: what came next.

No follow-on payloads have been observed. BleepingComputer reported that the script never stole credentials directly. That suggests the attackers intended to collect data by hand over the tunnel.[3][5] According to Ars Technica, extensions run with the editor's permissions. So this one likely had access to local files, source code and developer tooling.[6]

The Cleanup

Cursor says it pulled the build as soon as it detected it, blocked further distribution, then shut down the whole marketplace as a precaution. It audited access and says no other extensions were tampered with.[4] It rotated its keys, rebuilt the marketplace, and brought it back online.[4] Cursor also worked with the hosting provider to take the attacker's server offline. That blocked the connection to it and disrupted the attack.[4]

According to researcher Kevin Beaumont, Anthropic pushed version 2.1.314 several hours after the previous version appeared on Cursor.[6] Early reports pointed to Open VSX, but the other registries were not affected. Microsoft confirmed the malicious packages were never published through the Visual Studio Marketplace or npm, and Socket found no evidence the bad build became publicly available through Open VSX.[2][3][7]

What Defenders Should Change

If you were exposed, Cursor and Microsoft warn that uninstalling is not enough. The payload drops standalone scripts that keep running.[2][4] Taking down the attacker's server does not clean your device either. The keys, scripts and startup entries have to be removed by hand.[2] Microsoft points organisations using Cursor to Cursor's own security-hardening guidance. That includes restricting extensions to approved publishers and setting an installation cooldown, a delay before newly published extension versions install. Our inference: a cooldown longer than the reported window of roughly 11 to close to 12 hours would likely have held this build back until Cursor pulled it.[1][2][3][4]

Autopsy Verdict

Our read: publishing under the legitimate identity is how the tampered build was delivered. With auto-update on by default, as Phorion's chief technology officer described it, users could have received it automatically. How the attacker gained that publishing access is still unknown.[2][3][4][5][6] That leaves the verdict: what the companies state, and what is still open. What is stated: Anthropic says its systems were not compromised, and Cursor says the key, persistence and host report landed, and tells everyone to clean up. Independent analysis from Microsoft and other researchers describes what the loader attempts.[1][2][3][4][5]

Unclear: how the attacker got publishing access, whether Cursor itself was breached, and whether anyone actually logged in to a working tunnel.[1][3][4] Even the basics wobble. Reports put the window at roughly 11 to close to 12 hours. Cursor's notice names Windows and macOS. WIRED also lists Linux, and the analyzed code has a Linux branch.[1][3][4][8] The general lesson: a supply-chain attack, one that hides inside software you already rely on, does not have to arrive as a download you choose. It can come as an update to a tool you already trust.

This time, no source has named who was behind it, or explained how they came to publish under Anthropic's name.[2][3]

What to do

  • And keep secrets out of plain environment variables where you can, so a quick scan of your machine reveals as little as possible.

Open questions

  • How did the attacker get access to publish under Anthropic's official extension ID/account on Cursor's marketplace, and what level of access did they obtain?
  • Was Cursor's own infrastructure breached? Cursor has not explained this and did not answer WIRED's questions.
  • How many of the up to 5,000 affected users actually had a working SSH tunnel (for example, SSH already running or Cursor run with admin rights), and did the attacker log in to any of them?
  • Was there any hands-on follow-on activity, such as file transfers or credential theft, over the tunnels? None has been observed, but files moved after access would not show up in the static sample.
  • Which platforms were affected? Cursor's notice names Windows and macOS, WIRED mentions Linux, and the code includes a Linux branch.
  • How long was it live? Reports range from roughly 11 hours to close to 12 hours.
  • The incident title mentions Open VSX, but Socket, Microsoft, Anthropic, OpenSourceMalware and StepSecurity's correction all say the malicious build was distributed only through Cursor's marketplace.
  • Who is behind the attack? No source attributes it to a specific threat actor.

Sources

  1. [1]WIRED — A Supply Chain Attack Put Malicious Anthropic Claude Extensions on Thousands of Devices
  2. [2]Microsoft Threat Intelligence — Inside a malicious Claude Code-branded extension and its persistent SSH backdoor
  3. [3]The Hacker News — Official Claude Code VS Code Extension Compromised to Install Persistent SSH Backdoor
  4. [4]Cursor
  5. [5]BleepingComputer — Malicious Claude Code extension on Cursor planted SSH backdoors
  6. [6]Ars Technica — Anthropic VS code extension on Cursor backdoored for nearly 12 hours
  7. [7]Brandefense
  8. [8]Cyber Security News