Field intelligence for AI-first professionalsVol. II · Nº 195 · Monday, November 16, 2026
← Phantom Notes
Nº 193AI ToolsNovember 13, 20265 min read

Cursor Says Its Breach Began With a Key Fireworks Exposed. The Attacker Got In Through Old Coding Traces Cursor Kept for Training, Some Holding Users' Unredacted Credentials.

Cursor's security update of November 13 says the attacker behind October's fake Claude Code extension used an access key Cursor had issued to Fireworks, exposed by Fireworks, to copy a read-only bucket of coding agent traces, mostly from 2025 and from sessions outside Privacy Mode that Cursor was permitted to use for training. Some traces held unredacted credentials, which led into Cursor's SaaS tools, archived non-production repositories, support data, and the credential used to plant the malicious builds on the cache serving Cursor's VS Code marketplace. Cursor puts exposure at under 0.01% of customers for API credentials and under 0.05% of non-Privacy-Mode users for traces, says the incident is contained, and has moved all inference off Fireworks. Fireworks' own bulletin does not mention Cursor. Notion's advisory shows the same Fireworks incident reached Notion AI chat transcripts. We read Cursor's update, Fireworks' bulletin, and Notion's advisory.

By T.W. Ghost


Cursor published the root cause of October's breach today, November 13, 2026, "after an extensive investigation over the past two weeks." The attacker obtained an access key that Cursor had issued to Fireworks, its inference vendor, and that was, in Cursor's words, "exposed by Fireworks." With it they copied a read-only storage bucket Cursor used to share data with Fireworks. The bucket held limited coding agent traces, mostly from 2025, from sessions outside Privacy Mode that Cursor "was permitted to use for model training." Some of those traces contained unredacted credentials. From there the attacker reached select SaaS tools, archived non-production repositories, some with business contact information, limited support data and emails, usage metrics and contact details, summaries of some agent conversations, and credentials users had sent to Cursor support. One stolen credential placed the malicious Claude Code builds directly on the cache serving Cursor's VS Code marketplace on October 29, for about eleven hours, reaching at most 5,000 users who requested specific cached versions; Open VSX never carried the malware. Our October post covers that part.

The story for anyone who uses an AI coding tool is not the marketplace. It is the bucket. The data that let an attacker in was not live traffic; it was a copy of users' coding sessions, kept for training and shared with a vendor, carrying live secrets. Cursor says fewer than 0.01% of customers had API credentials exposed and fewer than 0.05% of non-Privacy-Mode users had traces exposed, all notified directly, with no base numbers given. It says the last attacker activity was November 1, that the incident is contained, that all inference has moved off Fireworks, that all internal credentials were rotated, secret scanning added, the marketplace rebuilt, and external forensics engaged. Fireworks' bulletin, last updated November 9, describes an LLMJacking-linked actor who reached Fireworks developer machines on October 24 through a third-party cloud provider incident the day before, and lists accounts, billing, secrets, datasets, and incomplete model weights as affected. It does not name Cursor or a Cursor key. We read Cursor's update, Fireworks' bulletin, and Notion's advisory. Here is what held up.


What Cursor Says Happened

Cursor's account (update dated November 13, 2026)
EntryA Cursor-issued access key "exposed by Fireworks"; attacker copied a read-only bucket used to share data with Fireworks
What was in the bucketLimited coding agent traces, mostly 2025, from non-Privacy-Mode sessions Cursor was permitted to use for training; some with unredacted credentials
ExposureAPI credentials for under 0.01% of customers; traces for under 0.05% of non-Privacy-Mode users; all notified (vendor numbers, no base)
Reached from thereSelect SaaS tools; archived non-production repos, some with business contact info; limited support data and emails; usage metrics and contact details; summaries of some agent conversations; credentials users sent to support
MarketplaceMalicious Claude Code builds placed on the cache serving Cursor's VS Code marketplace Oct 29 via a stolen credential; about 11 hours; at most 5,000 users; Open VSX never had the malware
Not affected, per CursorPrivacy Mode prompts and code; live traffic; production infrastructure; app updates and signing keys; Cloud Agents; Grok Bot and other SpaceX SI products; other extensions; Cursor Marketplace plugins; customer repos
StatusLast attacker activity Nov 1, blocked; Cursor "believe[s] the incident is contained"
RemediationAll inference moved off Fireworks; all internal credentials rotated; secret scanning added; marketplace rebuilt; external forensics engaged
DatesUpdate says Cursor "shared" on Oct 29 and "published a security notice" Oct 30; our October post dates the notice Oct 29 with an Oct 30 indicator update

The Other Two Accounts

DetailWho
Fireworks bulletin (updated Nov 9)LLMJacking-linked actor; third-party cloud provider incident Oct 23; Fireworks dev machines reached Oct 24; accounts, billing, secrets, datasets, incomplete model weights affected; Cursor not namedVerified (Fireworks)
Notion advisory (Nov 12, updated Nov 13)Under 1% of workspaces; Notion AI chat transcripts used for Insights, mostly unattributable; disconnected from Fireworks Nov 3; data deleted Nov 4; Insights disabled; exposed Notion API keys revoked Nov 11Verified (Notion)
Attacker SSH key and persistent backdoorPrior (Microsoft, Oct 30)

Training-data copies are attack surface. The traces that carried live credentials to a vendor bucket were kept because Cursor was allowed to train on them. Every tool that keeps sessions for training keeps this risk.

One-sided attribution. Cursor says a key Fireworks exposed. Fireworks' public bulletin describes its own incident and does not mention Cursor. Until Fireworks speaks to it, the chain has one author.

Percentages without a base. "Under 0.01%" and "under 0.05%" are Cursor's figures and come with no customer or user counts to multiply them by.

Privacy Mode was the line. Users in it were not in the bucket. Users outside it were. That is the whole product decision, in hindsight.


Verified vs Unconfirmed: The Scorecard

ClaimVerdict
Update dated Nov 13; two-week investigationVerified (Cursor)
Key exposed by Fireworks; bucket copied; traces with credentialsCursor's account
Under 0.01% and under 0.05%; all notifiedVendor numbers, no base
Downstream access list; marketplace mechanism; 5,000 users; 11 hoursCursor's account
Contained; last activity Nov 1Vendor statement
Not-affected list; remediationVendor statement
Fireworks timeline; Cursor not namedVerified (Fireworks bulletin)
Notion scope and datesVerified (Notion advisory)
Oct 29 vs Oct 30 notice dateMinor discrepancy; both kept as Cursor states them

What To Do

Turn on Privacy Mode if you have not, because it was the boundary between the users in the bucket and the users who were not.

Rotate any key you ever pasted into a Cursor agent session or sent to Cursor support, and still run the October 29 and 30 indicator checks; the root cause does not retire the backdoor.

The frontier is elsewhere: the same Fireworks incident reached Notion's AI transcripts, and the question for every vendor is now the one Cursor answered the hard way: what do you keep, where, and who else holds a key. Our model comparison has the current standings side by side.


Sources


*Not sure what your coding tool keeps from your sessions, or who it shares it with? Take the free 2-minute quiz and get matched. Then read about the backdoor this investigation explains: the Claude Code extension in Cursor's marketplace.*

Which model should you be using?

Three minutes, twelve questions, one defensible answer.

Take the quiz →