Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

Cursor traces a leaked key to a bad extension

On November 13, 2026, Cursor said a Fireworks-exposed key reached old agent traces and a malicious Claude Code extension cached for about 11 hours.

RA
· Founder, AIToolRanked
Four dark cards in a row on a black field, linked by arrows: a key, stacked bars, a canister, and an open lock
On this page

What Cursor reported on November 13

On November 13, 2026, Cursor published An update on our recent security incident ↗. The picture above is ours.

Fireworks exposed an access key Cursor had issued to it. The attacker copied a read-only bucket of limited coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which Cursor says it was allowed to use for model training. Some traces still held credentials. Cursor says less than 0.01% of customers had API credentials exposed. From those, Cursor says the attacker reached select SaaS tools and then: archived non-production source repositories, some with business contacts; limited support data and email; usage metrics and contact details; summaries of some agent conversations; and credentials users had sent to support.

On October 29 a stolen internal credential let the attacker put malicious copies of Anthropic's Claude Code extension on the cache for Cursor's VS Code marketplace. Cursor says Anthropic did not publish them and that this was not a flaw in Anthropic's product. Open VSX never had the malware. At most 5,000 users who requested specific cached versions that day potentially received the build, for approximately 11 hours. Cursor removed it, took the marketplace offline, and worked with the host to take the attacker's server down.

The security notice ↗ names Anthropic.claude-code 2.1.312 and 2.1.313 for Windows and MacOS. The page is dated October 29, 2026, and says: "Update: Oct 30, 2026, 11:30 am PT: We updated the IOCs and clarified affected versions." The November 13 post says the notice was published October 30. The window on the notice is about 2:27am PT to 1:35pm PT on October 29.

Microsoft ↗ (October 30) calls it a persistent SSH backdoor, says the packages were never on the Visual Studio Marketplace or in npm @anthropic-ai/claude-code, and says Anthropic confirmed its systems were not compromised. Cursor says the last activity it saw was November 1 and was blocked, and that it deactivated the Fireworks key.

Two counts, and what was spared

Under 0.01% of customers had API credentials in the traces. For users not in Privacy Mode, traces were exposed for fewer than 0.05% of users. Cursor says it notified each of them: "If you did not receive a notification, your data was not in these traces." The 5,000 is who potentially received a cached build, not the credential count.

Cursor says the set did not include Privacy Mode prompts and code, live traffic, production systems, official updates, release signing keys, Cloud Agents, Grok Bot, "all other SpaceX SI products," other extensions, active inference, or customer repositories. It moved inference off Fireworks.

@cursor_ai ↗ posted November 14 at 01:13 UTC: "We’ve published an update on our recent security incident. See a detailed overview of what happened and our remediation steps here."

Uber notices tied to Fireworks

TechCrunch ↗ (Zack Whittaker, November 13) reported Uber is notifying drivers that personal and driver's license information was stolen at a provider that told Uber on October 9. TechCrunch learned the provider is Fireworks AI. Uber's statement: "The notices to drivers are in relation to a security incident impacting one of our third-party vendors, Fireworks.ai, which had access to some Uber data."

Counts in notices TechCrunch saw: 2,567 in Massachusetts, 3,430 across South Carolina, 1,873 in Rhode Island. Illinois: number pending. California: no count. No total from Uber. TechCrunch says the number is expected to rise. Neither page puts those licenses in Cursor's bucket. Notion's part is the earlier Fireworks post.

Am I affected, and what to do

Cursor says it contacted each affected customer, and that most customers were not affected.

If you installed or updated Anthropic.claude-code 2.1.312 or 2.1.313 from Cursor's marketplace on macOS or Windows in that October 29 window, follow the security notice ↗. It says the package appends an SSH key labeled svc-telemetry, and that uninstalling "may not be sufficient" because persistence scripts keep running. Use the check and the script on that page. The indicator list also names Linux paths under ~/.config and a crontab entry. Write to security@cursor.com with questions.

The notice says that on most macOS and Windows hosts "the attacker could not have logged in through the tunnel," and that the extension still added the key, installed persistence, and sent host details. Microsoft says those writes can happen when the tunnel does not connect.

A direct notice covers traces and support data. Cursor says it emailed admins the keys to rotate, and notified customers if support or business data "may have been involved." No email clears the extension check. Copied traces came from sessions outside Privacy Mode. The post does not describe a way to pull those copies back.

What people are saying

  • "Any key you paste into an agent chat can end up in a log. Rotate it after." — Paul ↗ (@PaulShipsAI). That is his rule, not a Cursor measurement.

  • "The Cursor security incident shows why securing coding agent traces matters" — Daniel San ↗ (@dani_avila7)

  • "I checked my machine when the first notice dropped. If you updated the extension on Oct 29, check yours." — OLAITAN Michael Olanrewaju ↗ (@lanre_olat)

  • "Privacy Mode was not the only line that mattered in Cursor's incident." — DY ↗ (@DYoung_AI), who also wrote: "I have not audited any machine for this."

Reddit was not reachable for this story.

Our take

Run the notice's check if Claude Code updated inside Cursor on October 29. The versions, the clock, and the store are above. Uninstalling is where the page starts.

Rotate a key pasted into an agent chat. Cursor's notified credential set is under 0.01% of customers. Paul's point is the habit. Support credentials are the separate list Cursor says it notified.

Do not fold Uber's license counts into the 5,000. Those notices are another Fireworks customer. The extension figure is people who potentially received a cached build, for about 11 hours. For the products, see Copilot vs Cursor, the Claude Code review, and Cursor Projects, which this update does not change.

More from the news desk · About Rai Ansar