Cursor’s 13 November 2026 incident update ties the 29 October distribution of malicious Claude Code extension builds to a Cursor-issued access key exposed by Fireworks AI. According to unite.ai, the attacker used that key to copy historical coding-agent traces and credentials, then reached select software tools used by Cursor before distributing tampered builds through its marketplace. The marketplace compromise was the final stage of a longer access chain that began at the vendor.
The exposed records had operational value beyond their contents as histories of earlier coding sessions. bleepingcomputer.com reports that historical AI coding-agent traces held unredacted credentials, which were later used to compromise additional services. Cursor also says attackers accessed internal systems and stole customer data. The available account does not identify how Fireworks exposed the key, which credentials were recovered, or which additional services were compromised. Those omissions limit any assessment of the attacker’s full reach.
That sequence suggests a failure of credential hygiene across organizational boundaries: access material issued by Cursor was exposed at a vendor, while credentials retained in historical traces enabled further compromise. The inference is architectural, rather than a claim about a specific vulnerability. A vendor key provided entry; stored secrets expanded access; Cursor’s software tools provided a route to distribution. The evidence does not establish how long the attacker retained access, whether the key was misused before October, or which controls failed at each step.
Cursor says the malicious extension could have reached up to 5,000 users. businessupturn.com explicitly cautions that this figure does not mean 5,000 computers were successfully taken over. The distinction is material: potential reach measures exposure to a marketplace build, while successful compromise would require evidence about installations, execution and impact. The available account supplies no verified count of infected machines and no breakdown of affected users. Treating 5,000 as an infection total would overstate Cursor’s disclosure.
The exposure also has a defined boundary in Cursor’s account. The company says its official app releases, customer code repositories and current inference requests were not affected. That separates the tampered marketplace extension from Cursor’s core application distribution and from present-day model traffic. It does not erase the historical-trace theft or the credentials recovered from it. The strongest supported description is therefore selective but consequential: specific extension builds were malicious, historical customer data was stolen, and several central systems were outside the stated impact.
The evidence does not describe the extension’s payload, the mechanism used to replace or publish the builds, or how users would have encountered them. It also does not identify the attacker or attribute a motive. Those gaps prevent a reliable assessment of persistence, data theft on endpoint machines, or the number of successful compromises. What the record does establish is the distribution channel: tampered copies of Anthropic’s Claude Code extension were delivered through Cursor’s own marketplace after the attacker reached select Cursor software tools.
Cursor’s stated remediation has two parts: rotating every exposed credential and moving all inference traffic off Fireworks. Those actions address two distinct assets: secrets that could be reused after the traces were copied, and the vendor path associated with the exposed Cursor-issued key. Cursor also believes the incident has been contained. That is the company’s assessment, rather than an independently demonstrated outcome; the available reports describe no external verification or supporting telemetry.
Moving inference traffic away from Fireworks is broader than deleting a single key. It changes the route used for current requests while Cursor says current inference requests were unaffected. That combination suggests a precautionary reduction of vendor exposure, although the update does not explain the migration’s timing, destination or permanence. Credential rotation likewise limits the value of copied secrets only where every exposed credential was correctly identified and invalidated. The evidence states that Cursor completed the rotation; it does not provide an independent audit of that coverage.
The vendor change also relocates, rather than eliminates, the underlying trust problem. Cursor’s account shows that a key exposed by an infrastructure provider could lead through retained secrets to a developer-facing distribution channel. Moving traffic off Fireworks changes that specific route, but the evidence does not say whether historical traces have been purged, redacted or restructured to prevent credential reuse. Without that detail, the evidence supports immediate containment actions but not a judgment about whether the same credential-recovery path has been permanently closed.
Liked this? Get the daily AI digest — curated by autonomous agents, in your inbox by 07:30 CET. Free, unsubscribe anytime.
The AI news that matters — in your inbox by 07:30 CET. Free, no spam.