Skip to main content
All incident reports
Incident report

Cursor says stolen Fireworks AI key led to malicious Claude Code extension on its marketplace

Cursor has confirmed that an access key exposed by Fireworks AI let an attacker copy internal data holding unredacted API credentials, and then publish trojanised copies of Anthropic's Claude Code extension on Cursor's marketplace on 29 October. Up to 5,000 users may have received the malicious builds. This report covers what Cursor says was taken, who else was hit through Fireworks AI, and what remains unclear.

· First published

The short version

  • The breach. Cursor says an attacker used an access key it had issued to Fireworks AI, which Fireworks exposed, to copy data from a read-only storage bucket and then reach other Cursor systems. The bucket held historical coding-agent traces from users who had not enabled Privacy Mode.
  • The malicious extension. On 29 October the attacker used a stolen credential to place modified copies of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, on the cache serving Cursor's marketplace. They were available for about 11 hours and up to 5,000 users may have received them.
  • What was exposed. Fewer than 0.05% of users had coding-agent traces in the bucket and under 0.01% of customers had API credentials compromised, Cursor says. Data taken from Cursor's SaaS tools included archived non-production source code, support conversations, usage metrics and contact details.
  • Other victims. Uber and Notion were also affected through the same provider, according to the sources. Uber says a service provider notified it on 9 October that the provider's internal credentials had been compromised and used to access Uber's account with the provider, and that Uber's investigation determined on 13 October that the credentials had been used to access Uber data; TechCrunch reported the provider was Fireworks AI.
  • The response. Cursor deactivated the Fireworks key, rotated credentials, moved inference traffic off Fireworks and rebuilt its marketplace. It says no further malicious activity has been detected since 1 November.

Timeline

DateEvent
OctoberFireworks AI discloses several security breaches, including one involving exposed API keys and tokens and another in which attackers reached development machines through a compromised third-party cloud provider.
9 OctA service provider tells Uber that an unauthorised third party compromised its internal credentials and used them to access Uber's account with the provider.
13 OctUber's investigation determines that the unauthorised party used the compromised credentials to access Uber data.
29 OctThe attacker places malicious copies of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, on the cache serving Cursor's marketplace; they remain available for about 11 hours.
1 NovCursor says no additional malicious activity has been detected since this date.
14-15 NovCursor's incident report is covered by BleepingComputer and The Hacker News, setting out how the key was obtained and what was taken.

How the attack unfolded

Cursor says an attacker obtained an access key it had issued to Fireworks AI, a third-party provider used for model training and inference, and that Fireworks exposed the key. The key opened a read-only storage bucket Cursor used to share data with Fireworks. The bucket held historical coding-agent traces, mostly from 2025, from users who had not enabled Privacy Mode, a Cursor setting that opts users out of their data being used for training.

Some traces contained unredacted API credentials. Cursor says the attacker used them to reach selected software-as-a-service tools and take archived non-production source code repositories, limited customer support data and email conversations, product usage metrics and contact details, summaries of some coding-agent conversations, and credentials customers had sent to the support team.

Cursor's monitoring detected the activity, and it deactivated the Fireworks key, rotated credentials and temporarily disabled some internal and product services. By then the attacker held a credential that allowed it to publish extensions to Cursor's marketplace. On 29 October it placed the modified extension on the cache used to serve that marketplace, where it stayed about 11 hours before removal.

Who is affected and what the malware did

Cursor estimates that at most 5,000 users who requested the affected extension versions during the exposure window may have received the malicious builds, and says it notified them. The malware ran JavaScript that collected system information, installed an attacker-controlled SSH key and tried to enable SSH access and persistence. Researchers found coding errors that likely stopped it working on many Windows and macOS machines where SSH was not already enabled.

Fewer than 0.05% of users had coding-agent traces in the bucket, and fewer than 0.01% of customers had API credentials exposed, according to Cursor. It says users with Privacy Mode enabled, customer code repositories, its production infrastructure, its own application and official releases, and other extensions and plugins on the marketplace were not affected. Anthropic's legitimate extension and Open VSX were not compromised.

Wider fallout through Fireworks AI

BleepingComputer places the Cursor breach in a run of attacks on AI developer platforms that the security company Wiz calls Black October. Stolen credentials were used for unauthorised inference, meaning the running of AI models, and produced large bills: at Railway around $740,000 in fraudulent AI charges. Fireworks AI has also been affected, and other customers of the provider include Notion, which disclosed that AI chat transcripts from a small number of workspaces were exposed.

Uber is a further case. It said a service provider notified it on 9 October that internal credentials had been compromised and used to access Uber's account, and that its investigation on 13 October confirmed data had been reached. TechCrunch reported that the provider was Fireworks AI; Uber has not named it, and has not said how many people were affected. The sources do not say whether the credentials used against Uber are the same ones taken in the Cursor breach.

Cursor says it has moved all inference traffic away from Fireworks, deactivated the key issued to it, cut off its access to shared data sets, revoked and rotated exposed credentials, rebuilt the extension marketplace and redeployed production infrastructure. It also worked with the cloud provider hosting the attacker's server to take that server down.

What is not known yet

  • Who carried out the attack; no source attributes it to a named individual or group.
  • Whether the SSH backdoor worked on any developer machine, and whether any machine was accessed remotely through it.
  • How many customers in total had data copied from Cursor's software-as-a-service tools, beyond the percentage figures given for traces and credentials.
  • Whether the credentials used against Uber came from the Cursor breach or from a separate compromise at Fireworks AI; Uber has not named the provider.
  • What other coding-agent traces, besides those mostly from 2025, were held in the shared storage bucket.

What to do

If you use Cursor

  1. Check whether you installed or updated Anthropic's Claude Code extension from Cursor's marketplace on 29 October; the affected versions are 2.1.312 and 2.1.313.
  2. If you did, follow Cursor's remediation guidance and revoke any credentials that were shared with Cursor support or left in coding-agent sessions.
  3. Inspect developer machines for unauthorised SSH keys, enabled SSH access and persistence mechanisms.

If you use Fireworks AI

  1. Rotate any API keys or tokens you issued to or shared with Fireworks AI, which disclosed breaches involving exposed keys and tokens in October.
  2. Review inference usage and billing for activity you cannot account for, as stolen credentials have been used to run up large AI charges.

For everyone else

  1. Scope access keys given to third-party providers to read-only, single-purpose use, and deactivate them when the integration ends.
  2. Review what you store in buckets and data sets shared with vendors, since logs and transcripts can carry unredacted credentials.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.