Threat
Cursor Traces Malicious Claude Code Extension to Third-Party Breach
First reported darkreading.com
Page published · Page updated
Earliest dated coverage: 17 Nov 2026 · First observed: 17 Nov 2026 · Latest dated coverage: 17 Nov 2026
Coverage timeline
Single-source incident — one report is available.
Why it matters
The Cursor compromise demonstrates how a breach at an AI infrastructure provider can cascade into malicious software delivery through a trusted coding-assistant marketplace.
Cursor traced malicious Claude Code extension builds in its marketplace cache to a breach at AI infrastructure provider Fireworks, involving a compromised key used to access Cursor data and credentials. Up to 5,000 users may have downloaded tampered versions 2.1.312 or 2.1.313 during an approximately 11-hour window on Oct. 29; the extensions attempted to collect system information.
Summary
Cursor traced malicious Claude Code extension copies distributed through its marketplace cache to a cascading third-party compromise. According to the reported investigation, attackers obtained a Cursor-issued key exposed by Fireworks, copied historical coding-agent session records, recovered improperly redacted credentials, and subsequently used a compromised credential to tamper with Cursor's extension cache. Fireworks separately reported that a cloud provider hosting its R&D systems had suffered a security incident on October 23, 2026.[0]
The unauthorized extension copies were available for approximately 11 hours on October 29. At most 5,000 users requesting cached versions 2.1.312 or 2.1.313 potentially received a malicious build; this is an exposure estimate, not a confirmed infection count. The extensions attempted system-information collection and persistent remote access, including installation of an attacker-controlled SSH key. Cursor stated that Anthropic did not publish these copies, that no Anthropic product vulnerability was involved, and that Open VSX never hosted the malware.[0]
The intrusion also affected Cursor-held information and downstream SaaS access. Cursor estimated that API credentials belonging to fewer than 0.01% of customers were exposed and historical coding-agent records belonging to fewer than 0.05% of users were accessed. Most affected records were from 2025, involved sessions outside Privacy Mode, and were eligible for model-training use. Credentials recovered from records enabled access to selected tools exposing non-production repositories, support data, emails, usage metrics, and other information.[0]
Cursor removed the malicious extensions, temporarily disabled and audited the marketplace, rebuilt its marketplace infrastructure, and rotated credentials. It also terminated Fireworks' shared-dataset access, restricted communications between systems and providers, and expanded secret scanning. Fireworks associated its breach with financially motivated criminals monetizing unauthorized LLM access, but the supplied report names no actor; all incident details in this dossier derive from one news report relaying the companies' statements.[0]
Attack chain
- 1. Upstream R&D environment compromise: Fireworks reported that a cloud provider hosting its R&D systems experienced a security incident on October 23, 2026. Attackers used access to those systems to obtain sensitive information, including API keys and access tokens. The report does not describe the initial intrusion technique.[0]
- 2. Cursor storage access through an exposed supplier key: According to Cursor, an attacker obtained an access key that Cursor had issued to Fireworks and that Fireworks exposed. The key allowed the attacker to copy historical coding-agent session records from a read-only storage bucket.[0]
- 3. Credential recovery and downstream SaaS access: Some copied session records contained credentials that had not been properly redacted. The attacker used these credentials to access selected SaaS tools containing non-production source repositories, customer support data, emails, usage metrics, and other data.[0]
- 4. Extension-cache poisoning: On October 29, the attacker used one compromised credential to tamper with the cache serving Cursor's VS Code extension marketplace and inserted unauthorized copies of Anthropic's Claude Code extension.[0]
- 5. Delivery through affected cached versions: The malicious copies remained available for approximately 11 hours. Users requesting cached versions 2.1.312 or 2.1.313 during that window could have received the tampered builds; Cursor placed the potentially exposed population at no more than 5,000 users.[0]
- 6. Attempted collection and persistent remote access: The malicious extensions attempted to collect information about affected systems and establish persistent remote access, including installation of an attacker-controlled SSH key. The report does not quantify successful execution or persistence.[0]
Disclosure timeline
| Date | Event |
|---|---|
| 2026-10-23 | A cloud provider hosting Fireworks' R&D systems experienced a security incident, according to Fireworks' subsequent account. Attackers obtained sensitive information, including API keys and access tokens.[0] |
| 2026-10-29 | An attacker tampered with Cursor's extension-delivery cache. Malicious Claude Code extension copies associated with cached versions 2.1.312 and 2.1.313 were available for approximately 11 hours.[0] |
| 2026-11-09 | Fireworks published an update describing the upstream hosting-provider incident and linking its breach to financially motivated criminals seeking to monetize unauthorized LLM access.[0] |
| 2026-11-13 | Cursor published expanded investigation findings connecting the extension incident to the key exposed by Fireworks, historical session records containing credentials, and subsequent credential abuse. The update also described estimated exposure and security changes.[0] |
How it works
The central mechanism was credential-mediated expansion of access rather than a disclosed software flaw. A supplier-held key provided read-only access to historical session records, but secrets retained inside those records enabled access to additional SaaS tools. Read-only permissions on the original storage location therefore did not contain the downstream consequences of the exposed data.[0]
The software-delivery compromise occurred in Cursor's extension cache. A compromised credential enabled unauthorized extension copies to be substituted into the delivery path for two cached versions. Cursor explicitly distinguished this cache tampering from an Anthropic product vulnerability and stated that the upstream Open VSX marketplace never contained the malware.[0]
The malicious extension behavior included attempted system-information collection and persistent remote access through an attacker-controlled SSH key. The supplied report does not provide payload code, execution triggers, installation paths, the key value, or confirmed persistence counts.[0]
Affected versions and patch status
| Product | Affected | Patch status |
|---|---|---|
| Claude Code extension copies delivered through Cursor's extension marketplace cache | Cached versions 2.1.312 and 2.1.313 requested during the approximately 11-hour exposure window on October 29, 2026. At most 5,000 users potentially received a malicious build; the report does not establish that every copy of these versions outside that delivery context was malicious. | Cursor removed the malicious copies, temporarily disabled and audited the marketplace, rotated keys, and rebuilt the marketplace. No fixed extension version is identified. Cursor stated that this was not an Anthropic product vulnerability and that Open VSX never hosted the malware.[0] |
Key takeaways
- A read-only historical dataset became a stepping stone to broader access because its contents included usable credentials. The consequential boundary failure was not limited to the permissions of the original storage key.[0]
- Scope this as a Cursor cache-delivery compromise, not a general compromise of Anthropic's Claude Code product or Open VSX. The reported exposure is bounded by two cached versions and a specific delivery window.[0]
- Keep the reported impact measures separate: up to 5,000 potential malicious-build recipients, API credentials exposed for fewer than 0.01% of customers, and historical records accessed for fewer than 0.05% of users. None is a confirmed endpoint-infection count.[0]
- Attribution and indicator-based hunting remain constrained: the supplied report identifies no named actor and supplies no literal attacker domain, IP address, malware hash, or SSH-key value. Fireworks' financial-motivation assessment concerns its breach and should not be expanded into a named attribution for the entire chain.[0]
Defensive actions
- Prioritize exposure review for endpoints that requested Claude Code extension versions 2.1.312 or 2.1.313 through Cursor's cache during the October 29 exposure window.: The reported delivery risk is specific to those cached versions and that approximately 11-hour window. Assess potentially affected endpoints for system-information collection and unauthorized SSH-key persistence, the behaviors attributed to the malicious builds; potential receipt alone does not establish infection.[0]
- Revoke and replace potentially exposed credentials, including credentials recoverable from affected historical records, and rotate affected internal credentials.: The attacker progressed from an exposed supplier key to credentials embedded in session records and then to SaaS and extension-cache access. Cursor reported credential revocation, replacement, and internal rotation as response measures.[0]
- Remove compromised supplier access to shared datasets and tighten communication permissions between internal systems and third-party providers.: Cursor terminated Fireworks' access to shared datasets and restricted service communications after the supplier-held key enabled access to historical records.[0]
- Extend secret scanning to retained coding-agent records and other internal data as well as source repositories.: Improperly redacted credentials in historical session records enabled further access. Cursor reported expanding automated secret scanning across repositories and internal data in response.[0]
- Contain and audit the compromised extension-delivery cache, remove unauthorized builds, and rebuild affected marketplace infrastructure before restoring trust in delivery.: The malicious copies were inserted into Cursor's cache rather than the upstream Open VSX marketplace. Cursor's response included removing the copies, temporarily shutting down the marketplace, auditing for further tampering, and rebuilding it.[0]