Research · AI security · Supply chain · AI coding agents

The Cursor breach: a vendor key, agent traces and a backdoored extension

How a leaked vendor key, stored AI coding agent traces and a cached extension turned into an SSH backdoor on developer machines, and what to check on yours.

7 min read

Short answer

Cursor says an attacker used an access key exposed by its vendor Fireworks AI to copy a shared bucket of old coding agent traces, some holding credentials that had escaped redaction. A second stolen credential let the attacker plant tampered copies of the Claude Code extension (2.1.312 and 2.1.313) in Cursor's extension cache on 29 October, which at most 5,000 users may have received. Anything you type into an AI coding agent can end up stored somewhere, so keep secrets out of agent sessions and treat editor extensions like dependencies.

Cursor has traced its October breach to an access key exposed by Fireworks AI, one of its vendors. The attacker used that key to copy stored coding agent traces, found credentials inside them, and used one of those to plant a backdoored copy of the Claude Code extension where Cursor users download extensions.

If you build with AI coding tools, two parts of this story apply to you whatever editor you use. Your agent sessions are data that someone stores, and your editor extensions are code that runs with your access.

Our sources are Cursor’s 13 November incident update and published analyses of the extension. We read them as reported by Cyber Security News and GBHackers. Cursor’s own page would not open for us, so every fact here appears in at least two of their reports. Cursor says the incident is contained.

What happened

Per Cursor’s update, the key gave read-only access to a storage bucket shared with Fireworks. The bucket held coding agent traces, mostly from 2025, from sessions that ran outside Privacy Mode. Some of those traces contained credentials that had escaped redaction.

A trace is a record of what an agent saw and did in a session: prompts, files, tool output. If a secret passed through the session, it can sit in the trace long after you’ve forgotten pasting it.

The attacker then used a credential taken from that data to reach Cursor’s extension delivery. On 29 October, modified copies of Anthropic’s Claude Code extension, versions 2.1.312 and 2.1.313, were placed in the cache that serves Cursor’s extension marketplace. Cursor says at most 5,000 users who asked for those specific cached versions may have received them, during a window of roughly 11 hours. Cyber Security News points out that this is a possible distribution count, not 5,000 confirmed infections.

Anthropic did not publish these packages, and Cursor says Anthropic’s product was not vulnerable. Some early reports said the bad package was on Open VSX. Cursor’s update says Open VSX never hosted it.

How one vendor key reached developer laptops

1. Vendor keyAn access key exposed by Fireworks AI, read-only
third party
2. Shared bucketOld coding agent traces from sessions outside Privacy Mode
stored sessions
3. Secrets in tracesSome credentials had escaped redaction
leaked secrets
4. Extension cacheTampered Claude Code 2.1.312 and 2.1.313 placed on 29 October
supply chain
5. Your laptopSSH key added, reverse tunnel attempted, env variable names sent out
backdoor
Our layout of the chain described in Cursor's 13 November update, as reported by Cyber Security News and GBHackers.

What the tampered extension did

The reports describe a loader added to the extension. On a machine that ran it, the code added an attacker’s SSH public key to the user’s authorized keys with the label svc-telemetry, then tried to open a reverse SSH tunnel to 134.209.82.225. It also sent out the names of environment variables that looked like credentials. Only the names went out, according to the reports.

The persistence files sit outside the extension’s own folder. So uninstalling or updating the extension does not remove them. Cursor’s cleanup advice, as reported, goes further than uninstalling for that reason.

There is one piece of luck in the reports: coding errors in the malware and missing administrator rights blocked remote login on most hosts. “Most” still leaves some.

The numbers Cursor gave

≤5,000users may have received a tampered buildpossible, not confirmed
~11 hwindow the tampered builds were served29 October
<0.05%of users had old agent traces exposedsessions outside Privacy Mode
<0.01%of customers had API credentials exposedfound inside traces
Source: Cursor's 13 November 2026 update, as reported by Cyber Security News and GBHackers. The figures describe different parts of the incident, so don't add them up.

The timeline

Attacker steps and Cursor's response

1Vendor key used to copy the shared trace bucketdate not given in the reportsattacker
2Tampered Claude Code builds placed in the extension cache29 October, about 11 hoursattacker
3Marketplace taken offline, bad builds removed, service rebuiltafter the builds were foundresponse
4Last attacker activity seen, and blocked1 Novemberresponse
5Update links the breach to the Fireworks key; incident called contained13 Novemberupdate
Source: Cursor's statements as reported by Cyber Security News and GBHackers. Red is the attacker, green is the response.

Per Cursor, Privacy Mode prompts, customer code repositories and official Cursor app updates were not affected.

Why builders should care even if they don’t use Cursor

The attacker never broke Cursor’s front door. They came in through a vendor, and the valuable thing they found was a pile of agent sessions with secrets in them. Most AI coding tools send your context to servers you don’t run: the tool maker’s, a model provider’s, sometimes a training pipeline. Each of those copies is a place a key can leak from.

The second half is an old problem in a new spot. We covered how fast a leaked key gets used in Leaked an AWS key on GitHub?, and how agents get steered by what they read in 6 real AI agent hijacks. Here the agent wasn’t tricked at all. The attacker swapped the tool itself.

What to do this week

If you use Cursor with the Claude Code extension. Check whether you had version 2.1.312 or 2.1.313 installed around 29 October. Look in ~/.ssh/authorized_keys for a key labeled svc-telemetry, and search your network logs for 134.209.82.225. If you find either, follow Cursor’s cleanup notice, isolate the machine, and rotate every credential it could reach from a different, clean device.

Keep secrets out of agent sessions. Don’t paste keys into prompts. Keep .env files and credential folders out of what the agent can read, and use short-lived tokens so a key copied from a trace six months later has already expired. If your tool has a privacy or no-training mode, find out what it changes about storage and turn it on for work code.

Rotate anything that ever passed through a session. If you can’t say for sure that a key never appeared in an agent chat, treat it as exposed and rotate it.

Treat editor extensions like dependencies. Know which marketplace your editor pulls from, pin versions for tools that run with your shell access, and give a brand-new release a day before you install it unless it fixes something you need.

Ask your AI vendors two questions. How long do you keep my session traces, and which of your own vendors can read them? This breach started two companies away from the people affected.

Sources

Checked on 17 November 2026. Cursor’s own update, Microsoft’s analysis and the Fireworks bulletin would not open for us, so we rely on these reports of them and only use facts that appear in at least two.

AI Security Engineer

Move into AI security from the role you have

340+ videos, 33 modules, labs on AWS, Azure and Google Cloud, a graded capstone and a certificate.

Launch price $199, regular price $299 one payment. Final price: nothing is added at checkout.

Enroll today See the full curriculum

AI Security Engineer

Enroll in the full course

A practical, applied AI security course. Every module, lab, quiz, handbook and the graded capstone open the day you join.

  • 340+videos
  • 33modules
  • 3clouds included
  • 340+ videos, plus every reading page and lab across all 6 parts
  • Local and cloud lab solutions for AWS, Azure and Google Cloud
  • Downloadable slide deck (PDF) for every module
  • Graded capstone and certificate
  • Every future update, including the quarterly future-watch

Launch price $199, regular price $299

Launch price until Nov 30, 2026One-time payment. No subscription. Final price: nothing is added at checkout.

You pay $199 USDUS dollars, paid by card. Final price: nothing is added at checkout.

You pay on Razorpay's secure payment page. We never see your card or bank details. One-time payment, no subscription.

  • Paid on Razorpay's secure page: Card, in US dollars
  • Refund policy