It started at Fireworks. Uninstalling Claude Code is not the close.
Cursor’s 13 November investigation says a key it issued to Fireworks, then exposed by Fireworks, opened a training-trace bucket, then SaaS, then the marketplace cache. Anthropic.claude-code 2.1.312 and 2.1.313 sat there about 11 hours on 29 October. Uninstalling the extension does not remove the SSH key.
Bottom line up front
CONFIRMED On 13 November 2026, after two weeks of investigation, Cursor published the missing middle of the 29 October marketplace incident. An access key Cursor had issued to Fireworks was exposed by Fireworks. The attacker used it to copy a read-only storage bucket of coding-agent traces (mostly 2025, sessions not in Privacy Mode). A small number of those traces held unredacted credentials. From there the attacker reached selected Cursor SaaS tools, then — on 29 October — used a stolen internal credential to plant unauthorized copies of Anthropic’s Claude Code extension directly on the cache that serves Cursor’s VS Code extension marketplace.
Do not collapse five facts.
- This is not an Anthropic product bug, and it is not Open VSX. Cursor: Anthropic did not publish the copy. The marketplace source, Open VSX, never had the malware. The implant sat in Cursor’s cache of
Anthropic.claude-code2.1.312 and 2.1.313 for Windows and macOS, about 2:27 a.m. to 1:35 p.m. PT on 29 October (~11 hours). Cursor’s 13 November update is the chain. The 29 October notice is still the hunt. - Uninstalling the extension is not the close. The 29 October notice: the payload drops standalone persistence (LaunchAgent
com.apple.sysmondon macOS; scheduled task SysMon / Startupsysmon.cmdon Windows), appends an SSH key labeledsvc-telemetry, and retries a reverse tunnel. Cursor’s own sentence on 13 November: uninstalling alone may not be enough. - “At most 5,000” is a request-window ceiling, not a census of implants. Cursor: at most 5,000 users who requested those specific cached versions on 29 October potentially received the build. That is not “5,000 confirmed backdoors.” Independent reporting (Phorion / OpenSourceMalware, via BleepingComputer 30 October) said coding errors likely blocked SSH enablement on many Windows and macOS hosts where SSH was not already on. The key, the persistence, and the beacon still landed. Hunt those, not the ceiling.
- Privacy Mode is not a hunt for the extension, and a missing email is not a hunt for the key. Cursor: Privacy Mode prompts and code were not in the traces. Non-Privacy-Mode traces: fewer than 0.05% of users, each notified. API credentials in those traces: less than 0.01% of customers. If you did not get a trace notice, Cursor says your data was not in that bucket. That does not answer whether a developer pulled 2.1.312/2.1.313 from the marketplace cache on 29 October.
- This is not yesterday’s Check Point, not N-central, and not a KEV row. Live catalog this run is still
2026.11.13/ 1781. [CVE-2026-123039](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/) is due tomorrow, 16 November. UnrealIRCd CVE-2026-114713 is the other Monday clock. Neither grew overnight. Cursor has no CVE and no KEV listing. Fireworks’s own October bulletins describe LLMJacking-shaped access to Fireworks secrets, billing, and some datasets — a vendor incident that Cursor now says supplied the key. Do not merge Notion’s AI chat transcripts, Railway’s fraudulent inference charges, or GTIG’s LLMJacking paper into this implant.
The operator sentence: if Cursor is on the estate and anyone could have updated Anthropic.claude-code from Cursor’s marketplace on 29 October, the floor today is hunt svc-telemetry and the persistence paths, rotate anything that host could reach, and treat uninstall as theater. A missing Cursor email is not a green row. Monday’s Check Point Jumbo is a different ticket.
What happened
15 November 2026, after yesterday’s 06:00 ET Daily Top. Yesterday’s live catalog was 2026.11.13 / 1781. It still is. CISA added nothing overnight. Yesterday’s Daily Top was the Check Point DynamicCLI KEV add. Cursor’s 13 November investigation write-up was not that story. The increment for this run is that investigation plus 14–15 November secondary reporting (BleepingComputer 14 November; The Hacker News 15 November) that publicly tied the 29 October marketplace notice to the Fireworks-exposed vendor key.
Cursor’s five-step picture, paraphrased from the 13 November post:
- It started at Fireworks. A Cursor-issued access key, exposed by Fireworks, copied a read-only bucket Cursor used to share data with the vendor. The bucket held limited coding-agent traces, mostly 2025, from sessions not in Privacy Mode that Cursor says it was permitted to use for model training. Some traces contained credentials that had not been redacted. Less than 0.01% of customers had API credentials exposed. From those credentials the attacker reached selected SaaS tools Cursor uses, then archived non-production source repositories (some with business contact information), limited customer-support data and email, product-usage metrics and contact details, summaries of some coding-agent conversations, and credentials users had sent to Cursor support.
- Cursor’s monitoring cut the key. Cursor deactivated the vendor key, opened IR, rotated credentials, and — out of caution — temporarily disabled internal and product services beyond what the evidence required.
- The attacker published a malicious extension. On 29 October, using a stolen credential from that internal exposure, the attacker placed unauthorized copies of Anthropic’s Claude Code extension on the cache that serves the Cursor VS Code extension marketplace. Cursor took it down, took the marketplace offline, and worked with the cloud provider hosting the attacker’s server to take that host down.
- Notifications. 30 October public notice with IOCs and remediation scripts; in-app and email to affected users and team admins; key lists to admins where credentials were exposed; direct notice where support or business data may have been involved.
- Containment, as Cursor states it. Last attacker activity seen 1 November, blocked. Inference moved off Fireworks. Marketplace rebuilt. Production infrastructure redeployed as a precaution. Signing keys and the production ship path: Cursor says the attacker never accessed them.
The 29 October notice is still the technical record of the payload. High-level and defensive, not a reproduction: on load, modified extension.js collects host/user metadata (hostname, username, OS, architecture, home path, environment-variable names, working directories, configured Git email, public IP); appends an attacker SSH public key commented svc-telemetry; tries to enable local SSH and open a reverse tunnel to 134.209.82.225; persists via LaunchAgent or scheduled task, retrying roughly every 60 seconds; beacons HTTP POST to that host on 8099. Cursor: turning on the local SSH service needs admin rights the extension does not have on most macOS and Windows hosts, so that step fails in those cases — the key, persistence, and host-detail POST can still succeed. Cursor: the attacker server no longer functions.
Fireworks’s own 9 November bulletin (update to a 9 October incident, plus a 23–24 October R&D-cloud hop) is a separate vendor narrative: Unit 42, LLMJacking-shaped financially motivated access, billing and some incomplete model weights/datasets for a limited set of accounts; Inference, Training SDK, and Nexus — ZDR by default — “not detected” as impacted. Cursor’s 13 November post is the first time Cursor names that Fireworks exposure as the door to Cursor. Do not treat Fireworks’s “as of 25 October, no further successful access on the customer platform” as a close of Cursor’s marketplace cache.
Not this increment as a reprint. [Yesterday’s Check Point DynamicCLI KEV](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/) has no catalog change (2026.11.13 / 1781 still). UnrealIRCd CVE-2026-114713 is the same Monday clock, still IRC. [N-central 4EverAble2Shell](/posts/2026-11-13-ncentral-4everable2shell-hotfix2/) and [ScreenConnect RMM+](/posts/2026-11-12-sckit-rmmplus-screenconnect/) are MSP consoles, already Daily Tops. SignDrop is a public local Windows PoC with no CVE and Microsoft saying it has not observed exploitation — not this toolchain. Craft CMS CVE-2026-123488 is reported probing, not KEV.
Why it matters
Three operational reasons this is the daily.
First: the false close is “we uninstalled Claude Code” or “we use Privacy Mode.” Privacy Mode answers the trace bucket. It does not answer the marketplace cache. The persistence Cursor named lives outside the extension directory. A developer laptop that auto-updated an official-looking Anthropic extension from the editor’s own marketplace is a trusted-channel problem, not a random VSIX sideload.
Second: the door was a vendor key Cursor issued, sitting where Fireworks could expose it, in front of a bucket that still held secrets. That is identity and data-handling, not a novel VS Code parser bug. Cursor rotated exposed credentials and asked issuers to revoke. Operators still have to assume any secret that appeared in a non-Privacy-Mode agent session, or that was pasted to Cursor support, may have been in that set — if Cursor emailed the admin. The 0.01% / 0.05% figures are vendor counts, not an inventory of your tenant.
Third: developer endpoints are production-adjacent. The payload’s job, as Cursor described it, was a persistent SSH path plus host reconnaissance (including environment-variable names). Independent researchers said the SSH-enable step was sloppy. Sloppy is not “safe.” A workstation that already had SSH on, or a later fix-up of the implant, is still a path to source, cloud CLIs, and the rest of the laptop’s token cache. This is not confirmed OT or ICS process impact. A plant whose engineers use Cursor still has an IT/developer incident until someone shows the same host writing a controller.
Who / what is affected
| Asset | What the sources say | Floor / action | Who should care first |
|---|---|---|---|
Cursor on Windows or macOS, Anthropic.claude-code 2.1.312 or 2.1.313 from Cursor’s VS Code marketplace, 29 Oct ~2:27 a.m.–1:35 p.m. PT | Cursor: at most 5,000 users who requested those cached versions potentially received the build | Hunt svc-telemetry and persistence; Cursor remediation scripts; isolate / rebuild on a hit; rotate from a clean host | Developer-workstation owners, CISO for AI coding tools |
| Same estate, Privacy Mode on | Traces not in the Fireworks bucket (Cursor). Marketplace cache is independent | Still hunt the 29 Oct window | Do not skip the extension hunt because Privacy Mode is default for teams |
| Cursor users not in that window / other extensions / Open VSX / Anthropic’s own channel | Cursor: other marketplace extensions validated; Open VSX never had it; Anthropic did not publish it | Out of the implant row; still rotate if Cursor emailed keys | Do not rebuild the fleet on this sentence alone |
| API keys / secrets in non-Privacy-Mode agent traces, or sent to Cursor support | Cursor: <0.01% customers for API creds; emailed specific keys to team admins | Rotate what Cursor named; hunt SaaS that those keys opened | Identity / cloud |
| Cursor production app, signing keys, customer code repos, live inference, Cloud Agents, Grok Bot, other SpaceX SI products | Cursor: not affected | Out of this row | Do not spend the hunt here |
| Fireworks customer accounts (billing, some datasets/weights, evaluation secrets) | Fireworks 9 Nov bulletin | Follow Fireworks notice; not this VSIX | Fireworks customers, including ones who are not on Cursor |
| Notion AI chat transcripts | Notion, via Fireworks, separate disclosure (REPORTED) | Notion’s own review flow | Notion admins — parallel vendor queue |
| Check Point Gaia Mobile Access; UnrealIRCd 6 linked nets | KEV due 16 Nov | [Yesterday’s Jumbo/hunt](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/); UnrealIRCd 6.2.7.1 | Perimeter / IRC — not this laptop |
No public census of how many of the “at most 5,000” actually executed persistence is in these sources. Treat the 29 October window as an inventory problem.
Technical context
High-level, defensive only. Not a reproduction.
Two stages, one vendor-named chain. Stage one is credential access against a Cursor-issued Fireworks key and a training-share bucket that still contained secrets. Stage two is supply-chain abuse of a marketplace cache, not of the upstream Open VSX registry and not of Anthropic’s publisher account as Cursor tells it. The cache is the lie: the editor thought it was serving Anthropic.claude-code.
The payload Cursor documented is persistence-first, not a credential stealer in the first stage. Environment-variable names and Git email go to the beacon. The SSH key comment svc-telemetry is the cheap host check. Persistence names impersonate Apple/Windows telemetry (com.apple.sysmond, SysMon). Cursor published SHA-256 for the modified VSIX builds (darwin-arm64 and win32-x64 for both 2.1.312 and 2.1.313) and for modified extension.js. Use those as supplementary — a rebuilt marketplace does not erase a laptop that already ran the loader.
Workaround is not a patch. There is no CVE. The close for the cache was Cursor taking the marketplace down and rebuilding it. The close for a host is hunt + key removal + persistence removal + credential rotation, and a rebuild if the SSH path or beacon evidence is there.
ATT&CK mapping for defender queries, not a playbook: T1195.002 Compromise Software Supply Chain (marketplace cache), T1552 Unsecured Credentials (traces), T1078 Valid Accounts (vendor key / stolen SaaS), T1098.004 SSH Authorized Keys (svc-telemetry), T1543.001 Launch Agent (com.apple.sysmond), T1547 Boot or Logon Autostart (SysMon / sysmon.cmd), T1572 Protocol Tunneling (reverse SSH to 134.209.82.225).
Public write-ups of the modified extension.js are untrusted input. They are not required to contain. Inventory the 29 October window, hunt the IOCs Cursor named, rotate, rebuild a hit. Do not replay them.
Threat / exploitation status
| Claim | Status | Confidence |
|---|---|---|
| Cursor 13 Nov: Fireworks-exposed Cursor-issued key → trace bucket → SaaS → marketplace cache | CONFIRMED as vendor investigation | High for Cursor’s statement |
Tampered Anthropic.claude-code 2.1.312/2.1.313 in Cursor marketplace ~11 hours on 29 Oct; SSH-backdoor attempt | CONFIRMED (Cursor 29–30 Oct notice) | High |
| At most 5,000 users who requested those cached versions potentially received the build | CONFIRMED as vendor ceiling | High for the statement; UNKNOWN for how many executed persistence |
| Anthropic product / Open VSX / Cursor production signing compromised | Cursor: no | High they are separate as stated; we are not independently auditing the registry |
| Privacy Mode traces not in the bucket; <0.05% users / <0.01% customers for traces / API creds | CONFIRMED as vendor counts | High for the statement |
| Last attacker activity 1 Nov, contained | CONFIRMED as Cursor’s assessment | Moderate — vendor IR, not a third-party close |
| SSH enable failed on many hosts without admin / without pre-existing SSH | REPORTED (Phorion / OpenSourceMalware via BleepingComputer 30 Oct); Cursor also says admin required | Moderate — still hunt keys/persistence |
| Fireworks October incidents are LLMJacking-shaped criminal access (Unit 42 / Fireworks) | CONFIRMED as Fireworks/GTIG context | High for Fireworks’s own incident; not a named actor on the Cursor implant |
| Same incident as Check Point 123039, N-central, ScreenConnect, Notion Insights | No evidence they are one operation | High they are separate queues |
| CISA KEV add overnight; CVE for this implant | No — catalog still 2026.11.13 / 1781 | High |
| OT/ICS process impact | UNKNOWN / not claimed | — |
| Ransomware use | UNKNOWN / not claimed | — |
What defenders should do
- Inventory the 29 October window before you trust a missing email. Which laptops run Cursor (Windows/macOS). Whether
Anthropic.claude-codecould have auto-updated from Cursor’s marketplace between 2:27 a.m. and 1:35 p.m. PT on 29 October. Extension version history, not “we use Privacy Mode.” Team admins: whether Cursor emailed a specific key list. Fireworks-direct customers are a parallel ticket.
- Hunt the artefacts Cursor named, then assume they will change.
authorized_keyscommentsvc-telemetry(macOS~/.ssh/authorized_keys; Windows user andC:\ProgramData\ssh\administrators_authorized_keys). Fingerprint Cursor published:SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk. Persistence:~/Library/LaunchAgents/com.apple.sysmond.plist;~/.config/.smt/.smtk; Windows%USERPROFILE%\.config\smt.cmd, task SysMon, Startupsysmon.cmd. Network (historical): 134.209.82.225, reverse SSHtunnel@that host TCP 22, HTTP POST:8099/beacon. Process:ssh -N -R '*:0:localhost:22',autossh -M 0,schtasks /create /tn SysMon. VSIX SHA-256 in the 29 October notice if you still have the package. Cursor’sgrep/Select-Stringchecks and remediation scripts are the vendor floor — a non-zero exit means something remains.
- A hit is a rebuild-class host, not an uninstall. Isolate. Remove the key, persistence, and tunnel processes. Rotate credentials and sessions from a trusted system — anything in the user token cache, cloud CLIs, Git, and secrets that host could see. Cursor: limited env-var collection does not rule out later access over a working backdoor. Do not use the suspect laptop as the log source of record.
- Treat the vendor key pattern as a control failure, not a one-off. Inventory inference/training vendors, the keys you issued to them, and any bucket you share for traces or fine-tunes. Secret-scan agent traces and support tickets. Privacy Mode (or equivalent no-train / ZDR) is a data-handling control; it is not a marketplace signature. Confirm Cursor’s rebuilt marketplace is what your fleet talks to.
- Keep the other clocks. [Check Point CVE-2026-123039](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/) and UnrealIRCd CVE-2026-114713 are due Monday, 16 November. A 0660 DynamicCLI socket is still not the implant close on Gaia. [Barracuda 116606/111560](/posts/2026-11-11-barracuda-esg-kev/) and Ivanti 118527 were due 13 November. [SonicWall CVE-2026-19868](/posts/2026-11-07-sonicwall-cve-2026-19868-kev/) is due 20 November. SignDrop and Craft 123488 stay on a watch list: public PoC / reported probing, not KEV this run. Do not publish or replay the modified extension.
RWP assessment
Score 8.9. It outranks a second day on [Check Point CVE-2026-123039](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/) (catalog unchanged; Monday clock already published), UnrealIRCd 114713 (same catalog, IRC, THN recap is not a new KEV row), SignDrop (public local PoC, Microsoft: no observed exploitation), Craft CMS CVE-2026-123488 (reported probing, not KEV), and the DOJ FaceStealer / Binance seizures (crypto-crime disruption, not this developer toolchain).
The material development is not “a bad VSIX showed up in October.” The 29 October notice left the how open. Overnight Cursor named a Fireworks-exposed vendor key, a trace bucket that still held secrets, a SaaS pivot, and a marketplace cache — and restated that uninstalling Claude Code is not the host close. We will not invent a 5,000-host implant census, an actor, a KEV row, ransomware, or OT process impact. We will not collapse Fireworks’s LLMJacking bulletin, Notion Insights, and this cache into one victim list.
What we are watching
- Whether Cursor revises the 5,000 ceiling, the 1 November containment date, or the “Open VSX never had it / Anthropic did not publish it” sentences.
- Independent IR that confirms
svc-telemetrypersistence outside Cursor’s notified set, or confirms the SSH-enable failures held. - Any CVE, KEV row, or federal clock for this cache — none this run.
- Fireworks follow-up that names the Cursor-issued key in Fireworks’s own bulletin, or expands the 23–24 October R&D-cloud hop.
- Check Point 123039 / UnrealIRCd 114713 through Monday 16 November; SonicWall 19868 toward 20 November.
- SignDrop only if Microsoft or CISA moves it from public PoC to exploited.
- Craft 123488 only if a vendor or CISA confirms successful compromise, not sensor hits.
Sources
- Cursor — An update on our recent security incident, 13 Nov 2026
- Cursor — Security notice — malicious Anthropic.claude-code 2.1.312 and 2.1.313, 29 Oct 2026 (updated 30 Oct)
- Fireworks AI — Update on October 2026 security incident, last updated 9 Nov 2026
- Fireworks AI — October 2026 security incident bulletin
- BleepingComputer — Cursor confirms security breach led to malicious Claude Code extension, 14 Nov 2026
- BleepingComputer — Malicious Claude Code extension on Cursor planted SSH backdoors, 30 Oct 2026
- The Hacker News — Cursor Says Stolen Fireworks AI Key Led to Trojanized Claude Code Extension Incident, 15 Nov 2026
- Google Threat Intelligence / Mandiant — The Rise of LLMJacking, 2 Nov 2026
- CISA — KEV JSON feed catalog 2026.11.13
- RWP Ventures — Check Point DynamicCLI KEV Daily Top 2026-11-14
- RWP Ventures — N-central 4EverAble2Shell Daily Top 2026-11-13
- RWP Ventures — ScreenConnect RMM+ Daily Top 2026-11-12
- RWP Ventures — LiteLLM cook.pth Daily Top 2026-10-25