Coding with AI

A Fake Claude Code Extension Attempted to Install a Backdoor on Your PC

Cursor warns fake Claude Code extension versions added an SSH backdoor; GPT-6 Cyber finds Linux and Firefox flaws with mixed independent tests.

Key Takeaways

  • Cursor warned that fake Claude Code extension versions 2.1.312 and 2.1.313 contained a malicious SSH backdoor that could add an attacker key, survive restarts, and report user details.
  • The fake extension targeted Windows and macOS for about 11 hours, but the SSH backdoor required administrator rights, reducing its impact on many machines.
  • OpenAI introduced GPT-6 Cyber, a security-focused model that finds software flaws and proves them with exploit code, with reported findings in Linux and Firefox.
  • Independent testing showed mixed results for GPT-6 Cyber, with strong performance in one benchmark but weaker results in another compared with OpenAI’s general-purpose GPT-6 Astra.
  • Developers should verify their Claude Code extension version, check SSH settings for the svc-telemetry key, use Cursor cleanup scripts, and audit extension stores used for AI coding tools.

Meta description: AI security trust issues surface as Cursor flags fake Claude Code extension versions and GPT-6 Cyber finds Linux and Firefox flaws under independent testing.

Tags: AI security, Claude Code, GPT-6 Cyber, Cursor, SSH

Featured image: abstract illustration of a code extension plug with a hidden lock and a shield scanning software flaws, friendly colors, no real people, no logos, no text.

For developers and security teams, the week of October 31-November 6, 2026 brought two concrete tests of AI security trust. The Real Python AI News weekly edition reported that Cursor warned about fake Claude Code extension versions 2.1.312 and 2.1.313, and that OpenAI described GPT-6 Cyber, a model for finding software flaws, with mixed independent test results. The edition also promoted a free five-minute developer roundup from the Real Python team and pointed to provenance for AI-written text.

Cursor Warns That Fake Claude Code Extension Versions Added an SSH Backdoor

Cursor, an app for writing code, published a security notice about two versions of the extension for Claude Code, Anthropic's AI coding assistant. The notice says versions 2.1.312 and 2.1.313 were tampered copies that Anthropic did not build. They were available for about 11 hours on October 29 and targeted Windows PCs and Macs.

The malicious copy focused on SSH, a standard way to log in to a computer from afar. An SSH key can act like a password-free pass for authorized access. The extension added an attacker's key labeled svc-telemetry to the computer's approved keys, started automatically when the computer restarted, and reported details such as the username. Cursor says the SSH-on step requires administrator rights, which causes it to fail on most machines.

After the warning, Cursor pulled both versions and got the attacker's server shut down. OpenSourceMalware said the copies from npm, Open VSX, and Microsoft's store are clean. Cursor users should check their installed version and look in SSH settings for a svc-telemetry key. Because uninstalling may not be enough, Cursor offers cleanup scripts.

Philipp Acsany of Real Python described the broader risk in the source edition: attackers like extension stores because one poisoned upload can reach everyone who trusts the listed name, and AI coding assistants are tempting disguises because they run close to company code and passwords. For companies whose staff code in Cursor, the check list is concrete: version number, October 29 time window, and svc-telemetry key name.

GPT-6 Cyber Finds Flaws, but Independent Scores Are Mixed

OpenAI introduced GPT-6 Cyber, a new AI model for security research. It hunts vulnerabilities, weaknesses in software that attackers could abuse, then proves each flaw is real by writing an exploit, working attack code that triggers it.

OpenAI says its researchers used GPT-6 Cyber to build exploits for 948 flaws in Linux, the software behind most of the world's servers. The company reports sending 304 critical findings to Linux's maintainers with suggested fixes. It also credits the model with finding 81 unknown high-severity bugs in Firefox, many already fixed.

Irregular, an independent testing company, saw a more mixed picture. The model cleared all 11 scenarios in one test set, but solved 90 of 288 challenges in another, slightly fewer than OpenAI's general-purpose GPT-6 Astra. Only vetted experts can use GPT-6 Cyber, according to OpenAI's description.

The model's value is not just the count of bugs found. It shows that a security-focused model can combine discovery with proof: it can identify a weakness and produce working attack code to demonstrate it. The mixed benchmark result also shows that one test set can look strong while another reveals limits.

What Developers Should Do With These Updates

For Cursor users, the immediate task is verification. Check whether the Claude Code extension is version 2.1.312 or 2.1.313. If it is, inspect SSH settings for svc-telemetry, use Cursor's cleanup scripts, and treat the machine as needing more than a normal uninstall. Teams should also audit extension stores used for AI coding tools because a trusted listing can become a delivery path.

For developers watching AI security models, GPT-6 Cyber is a sign that security research is becoming more tool-driven, but more specialized. The Linux, Firefox, and independent testing numbers give a basis for comparison: 948 exploit cases, 304 critical findings, 81 Firefox bugs, 11 clean scenarios in one set, and 90 of 288 challenges in another.

The Real Python AI News edition also mentioned provenance for AI-written text, but no product or method was detailed. That gap matters: as AI-written text becomes part of developer workflows, teams may need more than trust in a model name. They may need evidence of where generated text came from, who reviewed it, and which tools touched it.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.