Independent · Daily · Every regionLatest edition: 21 November 2026
AI WorldwireVerified AI news from every region Subscribe
Latest

Breaking Analysis

Cursor Breach Shows AI Key Theft Is a Supply Chain Risk

Cursor says an access key exposed at Fireworks AI led to a fake Claude Code extension. Google and Wiz describe a wider wave of AI credential theft.

One key exposed at an AI infrastructure vendor ended with a fake coding-assistant extension in a developer marketplace. Cursor's 13 November report is the clearest account yet of how that chain worked.

Key takeaways

What Cursor says happened

Cursor's incident update, published Friday 13 November, starts at Fireworks AI, a third-party provider it used. An access key Cursor had issued to Fireworks was exposed by Fireworks, and the attacker used it to copy data from a read-only bucket holding coding-agent traces, mostly from 2025, from sessions where users had not enabled Privacy Mode [1].

A small number of those traces contained credentials that had not been redacted. Cursor says fewer than 0.01% of customers had API credentials exposed and fewer than 0.05% of users had traces exposed [1]. Starting from those credentials, the attacker reached select software-as-a-service tools and, from them, archived non-production source code, limited support data, usage metrics and some credentials users had sent to support [1].

BleepingComputer reports that the poisoned extension tried to install persistent SSH backdoors, though researchers found coding errors that likely stopped it working on many Windows and macOS machines where SSH was not already enabled [2]. Cursor says its official app releases, customer code repositories and Privacy Mode users' code and prompts were not affected, and that the last attacker activity it saw was on 1 November and was blocked [1]. Cursor also says uninstalling the extension may not be enough for anyone who installed the tampered versions [1].

Why one key mattered

Fireworks' own update says it later found impact to account and billing information, and, for a limited set of accounts, to secrets, datasets and incomplete models. It says evidence indicates links to a financially motivated group aiming to sell tokens from unauthorised LLM access to illicit routing gateways [3]. Fireworks also reports a separate incident on 23 and 24 October, traced to a third-party cloud provider that hosted its research machines [3].

Google Threat Intelligence Group and Mandiant describe the business model. In an early-November post they report incidents in which attackers hijacked enterprise access to large language models, with fraudulent charges above $100,000 in a single day and cumulative charges reaching $1 million at some organisations before access was revoked. They also describe "LLM Mixers", gateways that proxy and resell unauthorised model access [5]. These figures are Google's, from its own incident-response cases.

Wiz's running tally lists at least nine AI developer platforms that disclosed incidents in October, plus more in early November, and says Fireworks, Zeabur and Lightning AI attribute their root causes to credential access [4]. Wiz also reports two cases where a stolen key was first tested within 30 minutes of the reported vendor breach, and one that then made roughly 800,000 model calls over thirty hours before it was remediated [4]. Wiz notes that most incidents were still being analysed and that reports on impact so far were mostly anecdotal [4].

Regional view

United States. Cursor's and Fireworks' own disclosures, read with Google's and Wiz's research, give the best documented link so far between a developer tool, a vendor and the LLMjacking market [1][3][4][5].

Singapore. The Cyber Security Agency of Singapore issued an advisory on 13 November saying the monthly average of reported incidents since September is three times the January to August average. It says the incidents do not appear to be coordinated campaigns against Singapore, cites compromised third-party suppliers and the growing use of AI by attackers, and urges phishing-resistant multi-factor authentication, least-privilege access and supplier oversight [6]. The advisory is general, not specific to AI platforms.

Taiwan context. Zeabur, a deployment platform that INSIDE describes as friendly to AI-assisted "vibe coding" developers, said in mid-October that it would shut down by 31 December after two security incidents [7]. Taiwan's INSIDE reports that the company is registered in the United States with no Taiwan subsidiary, and that its status page described access to a platform database and exposure of environment variables, account details and API tokens [8]. Wiz links the shutdown to the incident [4].

Our analysis

The notable fact is not that an extension was poisoned. It is where the credential came from: a vendor holding a copy of historical agent traces. AI coding tools generate logs and traces that often contain pasted secrets, and those logs are now shared with training and inference vendors. When a trace store leaks, it can behave like a password dump.

Limits apply: the Cursor account is the company's own, most of the wider campaign is still under analysis, and we have not seen independent confirmation of how many people installed the tampered builds beyond Cursor's "at most 5,000".

What developers and teams can check

What to watch next

FAQ

Was Anthropic's Claude Code hacked? Cursor says Anthropic did not publish the malicious copies and this was not a vulnerability in Anthropic's product; the copies were placed on Cursor's marketplace cache [1].

What is LLMjacking? Stealing or abusing access to an AI model account or API key, then selling or using it at the victim's expense [3][5].

Who should rotate credentials? Cursor says it emailed team admins a list of the specific keys to rotate, and that anyone who installed versions 2.1.312 or 2.1.313 should follow its security notice [1]. In our view, any secret pasted into an AI session whose traces a vendor stores is also worth rotating.

This article is analysis, not investment or legal advice.

Sources

  1. An update on our recent security incident, Cursor, 13 November 2026. Primary.
  2. Cursor confirms security breach led to malicious Claude Code extension, BleepingComputer, 14 November 2026.
  3. Fireworks AI security bulletin update, Fireworks AI, November 2026. Primary.
  4. Black October for AI developer platforms, Wiz, updated to 10 November 2026.
  5. The Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model Resources, Google Threat Intelligence Group and Mandiant, November 2026. Primary.
  6. Advisory on Recent Uptick in Reported Cyber Incidents, Cyber Security Agency of Singapore, 13 November 2026. Primary.
  7. Zeabur is shutting down: what happens next, Zeabur, October 2026. Primary.
  8. Zeabur 宣布停止營運, INSIDE (Taiwan), 17 October 2026.

AI Worldwire is independent and has no commercial relationship with the companies covered, unless stated. Research and drafting are assisted by AI; every article is reviewed and approved by a human editor before publication. Subscribe to the AI Worldwire newsletter.

Disclosure: the AI system that assisted this article, Claude, is made by Anthropic, one of the companies covered here.

Subscribe to the AI Worldwire newsletter