Supply Chain

Cursor breach: how a vendor's AI key became a backdoored Claude Code extension

A key Cursor issued to Fireworks AI exposed agent traces, leaked credentials and a marketplace cache. The attack chain, why it worked and what to change.

By · · 5 min read

On 13 November 2026 Cursor published a detailed update on the incident it first disclosed on 29 October. It began with an access key Cursor had issued to Fireworks AI, its training and inference provider, which Cursor says Fireworks exposed. A few steps later, malicious copies of Anthropic’s Claude Code extension were served from the cache behind Cursor’s VS Code extension marketplace for about 11 hours.

No step needed a zero-day. Each reused something that already existed: a key, a data copy, a plain-text credential and a distribution channel clients trusted by default.

The attack chain, step by step

  1. A vendor key is exposed. The Fireworks-held key could read a storage bucket Cursor used to share data with the vendor.
  2. Agent traces are copied in bulk. The attacker copied historical coding-agent traces from that read-only bucket. They came from sessions not in Privacy Mode, mostly from 2025, which Cursor was permitted to use for model training. Cursor says fewer than 0.05% of users had traces exposed.
  3. The traces contain secrets. Some traces held API credentials that had never been redacted. Cursor puts the share of customers with exposed API credentials below 0.01%.
  4. Pivot into internal SaaS. With those credentials the attacker reached SaaS tools used by Cursor. There they accessed archived non-production repositories, limited support data and emails, usage metrics and contact details, summaries of some agent conversations, and credentials users had sent to support.
  5. A marketplace credential is stolen. On 29 October the attacker used a credential taken from that internal data to place tampered copies of the Claude Code extension, versions 2.1.312 and 2.1.313, directly on the cache that serves Cursor’s VS Code extension marketplace.
  6. The payload runs on developer machines. According to BleepingComputer, the modified extension ran JavaScript that collected system information, installed an attacker-controlled SSH key, tried to enable SSH access and set up persistence. Researchers found coding errors that likely broke the backdoor on many Windows and macOS hosts where SSH was not already enabled.
  7. Exposure and containment. Cursor estimates that at most 5,000 users who requested those cached versions on 29 October may have received the malicious build. The last attacker activity it observed was on 1 November, and it was blocked.

Anthropic’s official extension was not compromised, and Open VSX, the upstream source for Cursor’s marketplace, never carried the malware. Cursor has moved all inference off Fireworks, rotated credentials and rebuilt its marketplace.

Not only Cursor

The Fireworks compromise reached other customers. Uber confirmed that thousands of drivers’ license records were stolen, and Notion confirmed that some AI chat transcripts were exposed. Wiz groups this wave of credential thefts against AI infrastructure as “Black October”.

Why it worked

Three assumptions failed at the same time.

“Read-only” was treated as “low risk”. A read-only bucket is harmless only if its contents are harmless. Agent traces are not. A coding agent sees environment variables, config files, terminal output and tokens pasted while debugging. A trace archive is a secrets store without a secrets store’s controls.

The data outlived its purpose. These traces were mostly from 2025. Every extra month of retention extended the window in which a leaked credential could still work.

The distribution layer was trusted implicitly. Clients installed whatever the marketplace cache served. Write access to the layer between publisher and user was enough. A cache that serves executable code is supply chain, whether or not anyone treats it that way.

A fourth point concerns the key itself. Cursor has now restricted where its credentials can be used from. A vendor key that works from anywhere is only as safe as the vendor’s weakest system.

What defenders should do

  • Redact secrets at source. Strip credentials from prompts, tool calls and outputs before traces are written, and scan existing trace stores.
  • Keep retention short. Give agent traces and training exports an explicit expiry. If nobody can name a use for a year-old trace, delete it.
  • Treat vendor inference and training keys as production credentials. Scope each to one bucket or prefix, restrict it to known networks, alert on unusual use and rotate it after any vendor incident.
  • Treat extension marketplaces and their caches as supply chain. Limit who can write to them, log every write and allow-list approved extensions.
  • Inventory where AI traces go. For each AI assistant in use, list which vendors receive prompts, code and traces, and for how long.
  • Enforce Privacy Mode or an equivalent opt-out at organisation level for tools that offer it, instead of leaving it to each developer.
  • Rotate any token that may have appeared in a prompt or tool call. Assume anything an agent read or printed is logged somewhere.

Checking whether you received the malicious versions

Look for Anthropic’s Claude Code extension at version 2.1.312 or 2.1.313. The editor’s extension list shows the version, and the folder name on disk includes it (for example anthropic.claude-code-2.1.312 in the Cursor extensions directory of the user profile). If either version was installed from Cursor’s marketplace during the exposure window, follow the steps, indicators and check scripts in Cursor’s security notice of 30 October. Cursor warns that uninstalling alone may not be enough.

Given the payload’s reported behaviour, also review ~/.ssh/authorized_keys for keys you do not recognise, check whether SSH or remote login was recently enabled, and look for new launch agents, scheduled tasks, systemd units or cron jobs. Then rotate every credential that was on that machine.

My take

The malicious extension is the visible part, but the decisive failure came earlier. Secrets were allowed to land in training data, and that data sat at a vendor for a long time behind a single key that worked from outside its intended location. After that, the rest was routine lateral movement.

Nothing about this pattern is specific to Cursor. Any company that ships agent traces to a model provider for evaluation or fine-tuning has built the same pipeline. Three questions are worth asking this week: where do our agent traces go, who holds a key to them, and would we notice if that key were used from an unfamiliar network?

Sources

  1. An update on our recent security incident · Cursor, 13 November 2026
  2. Cursor confirms security breach led to malicious Claude Code extension · BleepingComputer, 14 November 2026
  3. Uber confirms hackers stole thousands of drivers' license information in data breach · TechCrunch, 13 November 2026
  • AI Security
  • cursor
  • fireworks-ai
  • claude-code
  • vs-code-extensions
  • secrets-management