Security

Cursor's trojanized Claude Code extension hit up to 5,000 users; OpenAI cuts Cursor's models November 12

A trojanized Claude Code extension distributed through Cursor's own marketplace hit up to 5,000 users, and OpenAI pulls the editor's models on November 12, citing distrust of owner SpaceX. The $60 billion all-stock deal now sits on Space Exploration Technologies shareholders' books.

Vincent JiangVincent Jiang · 4 min read
Share
A developer works across panels of source code, pull requests and terminal output, the environment where a trojanized VS Code extension surfaced
1 / 5Slide 1 of 5
A developer's screen with source code and pull requests: the environment Cursor says was exposed when a trojanized Claude Code extension appeared in its marketplace. (Courtesy of Anysphere)

Cursor, Anysphere's$30B — Anysphere, private, latest valuation $30B AI code editor and SpaceX's−3.49% — SpaceX, down 3.49 percent today $60 billion acquisition, disclosed on October 30 that a modified, compromised Claude Code extension for VS Code was distributed through its own extension marketplace, affecting at most 5,000 users 12. The company pulled the marketplace, worked with providers to remove the malware and reinstated it with new guidance, telling users to update Claude Code for VS Code to version 2.1.314 1. The tampered code was not written by Anthropic$350B — Anthropic, private, latest valuation $350B 1.

A backdoor hidden in a trusted update

The details sharpen the risk. The malicious build, version 2.1.313, appeared under Anthropic's legitimate extension ID and carried a 66-line loader that added an attacker-controlled SSH key, installed persistence and sent host data to a remote server 5.

The captured material included the names of environment variables containing keys, tokens or secrets 5. Cursor's analysis puts the exposure window at roughly 2:27 a.m. to 1:35 p.m 5. Pacific on October 29 5. Three days earlier, JFrog's security researchers had reported 42 hijacked GitHub repositories carrying payloads that fire when developers open folders in AI coding tools, Cursor among them 6. What is new now is a confirmed, company-disclosed incident inside Cursor's own marketplace, not another research finding.

The buyer of record: Space Exploration Technologies (SPCX) shareholders

SpaceX paid $60 billion in stock for Cursor, closing on August 14 and folding the editor into SpaceXAI, SpaceX's AI arm, on about $2.6 billion of annualized B2B revenue, more than 50,000 customer teams and the majority of the Fortune 500 3. The purchase price rests on teams keeping their subscriptions, and the buyers include Salesforce+0.72% — Salesforce, up 0.72 percent today, which reports developer velocity up more than 30%, and NVIDIA−3.36% — NVIDIA, down 3.36 percent today, whose 30,000 developers commit three times more code with Cursor 3.

Space Exploration Technologies (SPCX) trades at a $1.95 trillion market cap, 71.06 times sales near a one-year low against a 101.8 historical median, with negative trailing EPS 1. Insider activity is one-sided: no insider purchases and $53.7 million of insider sales in the past year 1.

The menu shrinks while the marketplace leaks

The breach lands in a month already crowded with trust questions for the company. OpenAI$500B — OpenAI, private, latest valuation $500B stops supplying models on November 12, stating: "We are making this choice because we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts" 4. Chief executive Michael Truell counters that OpenAI models serve about 5% of Cursor traffic 4.

Still, bundled OpenAI quotas vanish from Premium tiers, third-party models bill from the first token after the cutoff, and the $8 X Subscription Plus tier is Grok-only 4.

Cursor trails every rival in reported developer usage

0%20%40%60%80%Claude Code65.5%GitHub Copilot58.7%OpenAI Codex29.5%Cursor25.5%
Data
Value
Claude Code65.5%
GitHub Copilot58.7%
OpenAI Codex29.5%
Cursor25.5%
Reported usage among respondents who use a listed AI coding agent, 2026 Stack Overflow Developer Survey, 30,903 valid responses across 169 countries; respondents could report multiple tools, so shares overlap. Source: [4].4

That chart is the pressure in numbers. A controlled test ran Cursor and Claude Code on the same Claude Sonnet 5.5 model against 186 hidden checks: both passed every run, seconds apart 4. When the editor itself stops separating the tools, buyers choose on which models are bundled and how predictable the bill is, and Cursor's bundled menu is thinning while its marketplace carried malware 4.

What Cursor's account limits

Cursor says the malicious build is removed, other extensions in its marketplace were not affected, and it is contacting impacted users directly 15. The OpenAI cutoff touches about 5% of traffic by Truell's account 4, and Grok, bundled at $8, gains a distribution channel inside a widely used editor 4.

Claude models are among the frontier models developers can choose inside Cursor 3, and no source reports Anthropic restricting access. The unresolved problem is precedent: OpenAI has set the position that a model provider may cut a long-standing customer over who owns the parent 4.

Renewals are the next reading

Enterprise renewal commentary and marketplace guidance through November are the figures to watch. The $60 billion price holds only while the teams that measure their coding tools closely keep paying; a disclosed breach and a shrinking bundled menu give buyers reasons to pause 134. If renewals slow, the cost lands on SPCX shareholders who funded the deal in stock, and Claude Code, GitHub−1.51% — GitHub, down 1.51 percent today Copilot and Codex CLI are positioned to collect 134.

Get Security by email, free

The breaches, flaws and security vendors AI is putting to the test. In our daily email, every weekday.

Loading SPCX

Reader comments

0 comments

    More in Security

    All in this topic →

    Security

    Suspected Chinese attackers rooted Palo Alto Networks firewalls as the stock ran 14% on AI fear

    Dream Security's report, verified by Palo Alto's own Unit 42, puts suspected state hackers inside US government-affiliated firewalls one week after the stock's 14% run priced AI-found flaws as a bigger-budget story.

    Security

    GitLab stock hit its 52-week high the day Alpine's GitLab was rebuilt from scratch

    A CVSS 10.0 flaw GitLab patched on October 12 nearly let an attacker run code inside the server that builds Alpine Linux, because the fix ships inside GitLab's own Ruby, not Ruby itself. Self-managed operators absorb the cost, and GTLB touched its 52-week high the day it surfaced.

    Security

    JFrog's lab found 42 hijacked repos wired to fire in Claude and Cursor; its stock is up 61% in 2026

    JFrog's lab published three finds this week that all sit on AI tooling: credential-stealing LiteLLM wheels, 42 booby-trapped repos wired to Claude and Cursor, and an actively exploited Brotli flaw. The market has priced the fear; the November guide will show whether it converts to bookings.

    Security

    Cloudflare's own disclosure: 155 GLM-5.3 responses leaked other customers' data, and Z.AI's name travels

    Cloudflare disclosed that 155 GLM-5.3 responses leaked other customers' data, and the fault was Cloudflare's configuration and SGLang's engine, not the Chinese lab whose name sits in the model ID. Z.AI's shareholders wear the headline anyway.

    Security

    Two firewall vendors patch in parallel while attackers already hold the keys

    Palo Alto Networks traded up 3.24% on its no-exploitation reassurance the same weekend it confirmed a second critical PAN-OS flaw, and Fortinet reports Oct 28 with patch details withheld until November. Meanwhile Group-IB researchers have the receipts: attackers already used both vendors' devices to harvest 70,000 credentials.

    Security

    Cellebrite cut off Serbia in 2025. The labs say its tools ran through the 2026 election anyway

    Cellebrite publicly cut off Serbia in February 2025. Three forensic labs say its extraction tools were used there anyway, into 2026, and the repetition after an earlier cutoff in Russia reads as a claim about the business model, not one bad customer.

    Get the newsletter

    The day's AI and market news, in your inbox every weekday.