Monday, November 16, 2026
AI desk
/
/
Cursor Says Exposed Fireworks Key Led to Trojanized Claude Code Extension

Cursor Says Exposed Fireworks Key Led to Trojanized Claude Code Extension

Cursor says an exposed Fireworks key led to a storage-bucket breach tied to its earlier malicious Claude Code extension incident.
Last updated
November 16, 2026
9 min read
Fact-checked

Photo: TechJournal

Share

Quick Answer

Cursor says an exposed access key issued to Fireworks gave an attacker read access to limited coding-agent traces and was connected to its October malicious Claude Code extension incident. Cursor says fewer than 0.05% of users had traces exposed, but affected extension users should treat the event seriously because uninstalling alone may not remove persistence. Review Cursor’s notices, rotate exposed credentials, and investigate affected devices.

Key Takeaways

  • Cursor said Fireworks exposed an access key that Cursor had issued to the AI company.
  • The key gave read access to a storage bucket containing limited, mostly 2025 coding-agent traces from sessions outside Privacy Mode.
  • Cursor said traces from fewer than 0.05% of users were exposed, and fewer than 0.01% of customers had API credentials exposed.
  • Cursor linked the access-key exposure to its October marketplace compromise involving malicious Anthropic.claude-code builds.
  • Microsoft found that the malicious extension attempted to establish persistent SSH access and collect host details.

What happened in the Cursor security breach?

The Cursor security breach involved an attacker obtaining an access key that Cursor had issued to Fireworks, then using that key to read a Cursor storage bucket containing limited coding-agent traces. Cursor disclosed the access-key exposure on November 13, 2026, and said the traces were mostly from 2025 sessions that did not use Privacy Mode. Cursor’s November 13 security update describes the storage-bucket access and its connection to the earlier extension compromise.

The most important point is that the November disclosure concerns more than a generic vendor incident. Cursor said the attacker used the exposed key in the broader sequence of activity that included malicious copies of the Anthropic.claude-code extension in Cursor’s VS Code extension marketplace. The two events involved different systems and different data risks, but users who installed the affected extension need to consider both risks when assessing their exposure.

Cursor said fewer than 0.05% of users had coding-agent traces exposed through the storage bucket. Cursor also said fewer than 0.01% of customers had API credentials exposed in those traces. Those figures limit the apparent scope, but an API credential can be highly sensitive because it may provide access to a cloud service, development environment, or software workflow until the credential is revoked or rotated.

Users should not assume that a low percentage means no action is necessary. Review any direct notification from Cursor, identify whether you used coding-agent sessions outside Privacy Mode, and rotate credentials that may have appeared in prompts, configuration files, logs, or tool output. The practical response is to focus on the sensitivity of the information in a trace, not only the reported percentage of affected accounts.

How did the Fireworks access key expose Cursor data?

The Fireworks access key exposed Cursor data because the attacker obtained a credential Cursor had issued to Fireworks, and the credential had read access to a Cursor storage bucket. Cursor said the bucket contained limited coding-agent traces, primarily from 2025, for sessions that did not use Privacy Mode. An access key is a machine credential that allows a service or application to authenticate to another system, so its permissions determine what an attacker can reach after theft.

Fireworks separately said its investigation found impacts to account and billing information for a limited set of accounts, along with secrets, datasets, and incomplete models for some customers. Fireworks also said payment-card numbers were not stored or affected. Fireworks’ security incident update provides the company’s description of its own affected information.

The Cursor and Fireworks disclosures do not establish that every Fireworks customer or every Cursor user was affected. The available disclosures instead describe a stolen or exposed credential that enabled specific access, with Cursor limiting its statement to the storage bucket and traces it identified. That distinction matters because breach notices often cover separate datasets, systems, and populations under one broad incident timeline.

Organizations that use hosted AI coding tools should review where secrets are stored and how third-party credentials are scoped. A credential with read access can still expose source context, configuration details, and prompt content, even if it cannot change files or infrastructure. Security teams should also avoid placing long-lived secrets in prompts or logs whenever a safer secret-management path is available.

Which Cursor extension versions were malicious?

The malicious Cursor extension versions were Anthropic.claude-code 2.1.312 and 2.1.313 for macOS and Windows. Cursor said an attacker used a stolen credential on October 29, 2026, to place malicious copies of Anthropic.claude-code in the cache serving its VS Code extension marketplace. The malicious builds remained available for about 11 hours, and Cursor said at most 5,000 users may have received them.

Incident detailCursor’s descriptionWhy it matters
Affected extensionAnthropic.claude-code 2.1.312 and 2.1.313Users can compare installed or previously installed versions against the identified builds.
PlatformsmacOS and WindowsBoth desktop platform groups need to assess possible exposure.
Availability windowAbout 11 hours on October 29, 2026The limited window narrows the investigation period but does not eliminate device risk.
Potential recipientsAt most 5,000 usersThe number is an upper bound stated by Cursor, not confirmation that every recipient was compromised.
Last observed attacker activityBlocked by November 1, 2026Cursor said it believes the incident is contained, although affected users still need remediation.

The affected versions matter because malicious extensions can run within a developer’s editor environment, where they may access project context and local system resources. Developers who rely on extensions for coding assistance should also apply the same caution used after other software-supply-chain events, including incidents involving exposed AI chat transcripts, where apparently routine work data can contain sensitive information.

Cursor said the attacker’s last observed activity was blocked on November 1, 2026, and it believes the incident is contained. Containment describes the company’s assessment of attacker activity in its environment. Containment does not mean a potentially affected computer is automatically clean, particularly when the malicious code may have created persistence outside the extension itself.

Why is uninstalling the malicious extension not enough?

Uninstalling the malicious Cursor extension may not be enough because Cursor said removal alone may not remove persistence mechanisms created by the malicious build. Persistence is a technique that lets malicious software remain active or regain access after the original installer, application, or extension is removed. The warning changes the appropriate response from a simple extension cleanup to a device-level security review.

Microsoft said its analysis found that the malicious extension attempted to add an attacker-controlled SSH key, enable SSH, establish reverse forwarding, create persistence, and collect host details. Microsoft’s technical analysis explains that these behaviors could support remote access and continued control of an affected machine.

An SSH key can allow authenticated remote access without a traditional password, while reverse forwarding can create a connection from the affected device back to attacker infrastructure. Those capabilities are concerning for developers because a workstation can hold source code, environment files, cloud credentials, deployment settings, and private repositories. The reported behavior was an attempted capability described in Microsoft’s analysis, not proof that every user who received the build experienced every action.

Users who installed Anthropic.claude-code 2.1.312 or 2.1.313 should remove the extension, but they should not stop there. Preserve relevant logs if your organization has an incident-response process, disconnect the device from sensitive resources when appropriate, and have a qualified security team inspect it for unauthorized SSH configuration, persistence, and suspicious remote connections. Stop and contact your employer’s security team, a managed security provider, or device support if you cannot confidently assess the system.

What should affected Cursor users do now?

Affected Cursor users should treat the identified extension versions and exposed trace data as separate remediation paths. The extension incident creates a possible endpoint-compromise risk, while exposed traces create a possible secrets and privacy risk. The safest approach is to investigate both paths rather than assuming that deleting an extension addresses credentials that may have appeared in old coding-agent sessions.

  1. Check whether Anthropic.claude-code 2.1.312 or 2.1.313 was installed on a macOS or Windows device.
  2. Remove the affected extension and preserve relevant system or security logs before making broad changes, if your organization has incident-response requirements.
  3. Rotate API credentials, tokens, SSH keys, and other secrets that may have appeared in coding-agent traces outside Privacy Mode.
  4. Review devices for unauthorized SSH keys, unexpected SSH service changes, reverse-forwarding activity, and persistence mechanisms.
  5. Contact your security team or a qualified incident-response provider if the affected device accessed work repositories, production systems, customer data, or cloud infrastructure.

Back up essential work files before performing a device rebuild or other major system recovery. A rebuild can be appropriate when persistent compromise cannot be ruled out, but it can also remove evidence needed for an investigation. Users who need to recover files from a damaged or wiped system should understand the limits of data recovery software before writing new data to the drive.

The stop line is clear: do not attempt to manually remove unfamiliar SSH keys or persistence entries on a work-managed device without authorization. Incorrect changes can disrupt access controls or destroy evidence, while incomplete removal can leave an attacker’s access path intact. Escalate the device to the responsible security or IT team when sensitive data or administrative access was involved.

Does Privacy Mode reduce the risk from this incident?

Cursor said the exposed storage bucket contained traces from sessions that did not use Privacy Mode, so Privacy Mode appears relevant to the specific trace exposure Cursor described. The disclosure does not mean Privacy Mode protects against every kind of endpoint compromise, malicious extension, credential theft, or third-party breach. Privacy settings can reduce certain data-retention or logging exposure, but they cannot replace device security controls.

The distinction matters because the October extension incident occurred through a marketplace cache, while the November disclosure concerned read access to a storage bucket. A privacy setting may change whether certain session traces are retained or included in a dataset, but a malicious local extension can attempt actions on the device where it runs. Users need separate controls for data handling, extension trust, credentials, and endpoint monitoring.

Developers should review what they paste into AI coding tools, especially API keys, access tokens, private certificates, production configuration values, and customer data. The same principle applies to other AI services that can process sensitive financial or work information, including tools that offer spending analysis from wallet data. Convenience features can be useful, but users need to understand what information they authorize a service to process.

The practical response is to use Privacy Mode where it fits your organization’s policy, while also removing secrets from prompts and code comments before sharing them with any external service. Organizations should define approved tools, credential-scanning practices, and response procedures for employees who discover that sensitive material entered an AI workflow.

How should organizations respond to the Cursor security breach?

Organizations should respond to the Cursor security breach as a potential software supply-chain and credential-exposure event. Security teams should first identify devices that installed the affected Anthropic.claude-code versions, then determine whether those devices had access to source repositories, cloud consoles, production credentials, or sensitive customer data. This prioritization matters because the greatest risk depends on what the potentially affected endpoint could reach.

A structured review should separate 4 areas: endpoint persistence, unauthorized SSH access, exposed secrets in coding-agent traces, and downstream service access. The categories are related but require different evidence and owners. Endpoint teams can assess device changes, identity teams can rotate credentials, cloud teams can review access logs, and legal or privacy teams can evaluate notification duties where applicable.

Organizations should also verify extension provenance and tighten marketplace controls. Restricting unapproved extensions, requiring device management, and monitoring unusual SSH changes can reduce exposure from future marketplace compromises. Broader cybersecurity programs need this layered approach because software trust, identity controls, logging, and endpoint protection address different points in the attack path. A basic overview of cybersecurity controls can help teams map these responsibilities to prevention, detection, and response.

Cursor’s statement that it believes the incident is contained is useful context, but each organization must still determine whether its own systems show signs of impact. The most sensible approach is to document findings, rotate relevant secrets, preserve evidence, and seek professional incident-response assistance when affected devices had privileged access or suspicious activity cannot be explained.

FAQ

Was the Cursor security breach limited to the malicious Claude Code extension?

No, the Cursor security breach involved both malicious Claude Code extension builds and read access to a storage bucket through an exposed Fireworks access key. Cursor said the bucket contained limited, mostly 2025 coding-agent traces from sessions that did not use Privacy Mode.

Which Claude Code versions should Cursor users investigate?

Cursor users should investigate Anthropic.claude-code versions 2.1.312 and 2.1.313 on macOS and Windows. Cursor said those malicious builds were available through its VS Code extension marketplace cache for about 11 hours on October 29, 2026.

Did the Cursor breach expose payment-card numbers?

No, Fireworks said payment-card numbers were not stored or affected in its incident. Fireworks said a limited set of accounts had account and billing information affected, while some customers also had secrets, datasets, and incomplete models affected.

Should affected users rotate API credentials?

Yes, affected users should rotate API credentials that may have appeared in coding-agent traces, prompts, configuration files, or related logs. Cursor said fewer than 0.01% of customers had API credentials exposed, but any exposed credential should be treated as compromised until replaced.

Has Cursor contained the malicious extension incident?

Cursor said it blocked the last attacker activity it observed on November 1, 2026, and believes the incident is contained. Affected users still need to inspect devices because Cursor warned that uninstalling the malicious extension may not remove persistence mechanisms.

Share this guide
Facebook
X
LinkedIn
Written by
Priya Sharma is a cybersecurity analyst and tech writer who covers digital privacy, online safety, and creative technology tools. She holds a CompTIA Security+ certification and writes about making security accessible for non-technical audiences. She’s passionate about the intersection of AI and creative work.

In this article

The AI Brief

Guides like this, every Friday.

One email. No hype cycle.

Essential guides