{"schema":"sanctity-signal/0.2","record_type":"signal","id":"cursor-fireworks-key-breach-e093","domain":"security","drop_id":"signal-e093","occurred_at":"2026-11-13","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Cursor","action_type":"publication of a breach post-mortem tracing the intrusion to an access key the company had issued to its inference vendor","observed":"Cursor published \"Security incident update\" on November 13, 2026, with no time of day, after about 2 weeks of investigation. The chain: Fireworks AI exposed an access key Cursor had issued it; an attacker used that key to copy data from a read only storage bucket. The bucket held mostly 2025 coding agent traces from sessions not in Privacy Mode, and some traces contained unredacted credentials. The attacker used those credentials to reach selected software as a service tools, and from there archived non production source repositories, limited customer support data and emails, product usage metrics and contact details, summaries of some coding agent conversations, and credentials users had sent to support. Separately, on October 29 the attacker used a stolen credential to place a malicious copy of the Anthropic Claude Code extension on the cache serving Cursor's VS Code extension marketplace. The tampered versions were 2.1.312 and 2.1.313, available for about 11 hours. Cursor states \"Anthropic did not publish this copy, and this was not a vulnerability in Anthropic's product,\" and that Open VSX, the marketplace source, never carried the malware. Cursor warns that \"Uninstalling the extension alone may not be enough.\" Scale: up to 5,000 users who requested specific cached versions on October 29 may have received the malicious build; fewer than 0.01% of customers had API credentials exposed; for users not in Privacy Mode, agent traces were exposed for fewer than 0.05% of users. Cursor says the incident \"started at Fireworks,\" moved all inference traffic off Fireworks, deactivated the key, cut its access to shared datasets, rotated credentials, took the marketplace offline, worked with the hosting provider to take down the attacker's server, and validated all other extensions as authentic. Dates: October 29, first notice and the malicious upload; October 30, indicators of compromise, remediation steps and notification of affected users and team admins; November 1, last attacker activity, blocked. Reported elsewhere and not in Cursor's post: Fireworks dates its own intrusion to October 24 and last unauthorized platform access to October 25, and its investigation later widened to billing data and some customer model files.","corrections":[],"evidence":{"status":"verified","note":"Verified at the primary: we read Cursor's post at cursor.com. It carries Friday November 13 and no time of day, which places it in Edition 091's or Edition 092's window. It was never carded, so it is carded here under its true event date, and we say plainly that it surfaced late.","sources":[{"role":"primary","label":"Cursor: Security incident update, cursor.com, Nov 13, 2026","url":"https://cursor.com/blog/security-incident-update"},{"role":"corrob","label":"Unite.AI: Cursor traces breach to access key exposed by Fireworks AI, Nov 13, 20:34 Eastern","url":"https://www.unite.ai/cursor-traces-breach-to-access-key-exposed-by-fireworks-ai/"},{"role":"corrob","label":"SecurityWeek: Malicious Claude Code extension served in Cursor supply chain attack","url":"https://www.securityweek.com/malicious-claude-code-extension-served-in-cursor-supply-chain-attack/"},{"role":"corrob","label":"Cybersecurity News: how a leaked Fireworks AI key led to a backdoored Claude Code extension","url":"https://cybersecuritynews.com/cursor-data-breach/"}]},"strength":{"cost":"Medium","commitment":"High","reversibility":"Low","persistence":"Years","rationale":"Cost is medium and real: naming your own vendor, publishing exposure percentages, and telling customers that uninstalling may not be enough all invite liability and lost deals. Commitment is high because a dated post-mortem with version numbers and named third parties cannot be walked back. Reversibility is low for the same reason. Persistence is years: the lasting item is not the incident but the finding that agent traces hold live credentials, which changes how every buyer writes a contract with an AI vendor."},"shift":{"prior":"AI supply chain risk meant your model vendor's own security.","action":"Cursor published \"Security incident update\" on November 13, 2026, with no time of day, after about 2 weeks of investigation. The chain: Fireworks AI exposed an access key Cursor had issued it; an attacker used that key to copy data from a read only storage bucket. The bucket held mostly 2025 coding agent traces from sessions not in Privacy Mode, and some traces contained unredacted credentials. The attacker used those credentials to reach selected software as a service tools, and from there archived non production source repositories, limited customer support data and emails, product usage metrics and contact details, summaries of some coding agent conversations, and credentials users had sent to support. Separately, on October 29 the attacker used a stolen credential to place a malicious copy of the Anthropic Claude Code extension on the cache serving Cursor's VS Code extension marketplace. The tampered versions were 2.1.312 and 2.1.313, available for about 11 hours. Cursor states \"Anthropic did not publish this copy, and this was not a vulnerability in Anthropic's product,\" and that Open VSX, the marketplace source, never carried the malware. Cursor warns that \"Uninstalling the extension alone may not be enough.\" Scale: up to 5,000 users who requested specific cached versions on October 29 may have received the malicious build; fewer than 0.01% of customers had API credentials exposed; for users not in Privacy Mode, agent traces were exposed for fewer than 0.05% of users. Cursor says the incident \"started at Fireworks,\" moved all inference traffic off Fireworks, deactivated the key, cut its access to shared datasets, rotated credentials, took the marketplace offline, worked with the hosting provider to take down the attacker's server, and validated all other extensions as authentic. Dates: October 29, first notice and the malicious upload; October 30, indicators of compromise, remediation steps and notification of affected users and team admins; November 1, last attacker activity, blocked. Reported elsewhere and not in Cursor's post: Fireworks dates its own intrusion to October 24 and last unauthorized platform access to October 25, and its investigation later widened to billing data and some customer model files.","update":"It means every key you ever issued one of them."},"inference":{"established":["Fireworks exposed a Cursor key; the attacker copied a read only bucket.","Tampered versions 2.1.312 and 2.1.313 were served for about 11 hours.","Cursor says Anthropic did not publish the copy and had no vulnerability."],"unknown":["How many users ran the backdoored build. 5,000 is the ceiling.","Who the attacker is. Cursor names no actor.","That the count is final. The Fireworks investigation widened later."]},"narrative":{"hook":"Cursor traced its breach to a key it gave an AI vendor.","plain":"On November 13 Cursor published its post-mortem: an attacker took an access key Cursor had issued to Fireworks AI. Cursor says the backdoored Claude Code extension it then served reached as many as 5,000 users.","story":["The key was read only. It held 2025 agent traces, and some traces contained credentials users had pasted in. Those credentials opened other systems. A coding agent's traces held live secrets, and nobody guarded them as secrets."],"trail":[{"d":"OCT 24","t":"An unauthorized party reaches Fireworks AI development machines; Fireworks detects the activity that day, per later reporting."},{"d":"OCT 29","t":"The attacker places a malicious copy of the Claude Code extension on the cache serving Cursor's extension marketplace. It is available about 11 hours. Cursor publishes its first notice."},{"d":"OCT 30","t":"Cursor publishes indicators of compromise and remediation steps and begins notifying affected users and team administrators."},{"d":"NOV 1","t":"Last attacker activity Cursor observes. It is blocked."},{"d":"NOV 13","t":"Cursor publishes the post-mortem naming the Fireworks key as the origin, the credentials inside agent traces as the escalation path, and up to 5,000 users as the exposure ceiling."},{"d":"Reacts","r":true,"t":"Fireworks AI, named as the origin by its own customer; Anthropic, whose extension name carried the backdoor; every company that issued an inference vendor a key to a bucket; VS Code marketplace operators and Open VSX; enterprise buyers asking what their coding agents logged; and auditors who will treat agent traces as a credential store."}],"glossary":[{"term":"Agent trace","def":"The stored log of what a coding agent read, wrote and ran. It can contain whatever the user pasted in, including live passwords and keys."},{"term":"Privacy Mode","def":"A Cursor setting that stops sessions being retained. Traces from sessions outside it were in the copied bucket."},{"term":"Inference vendor","def":"A company that runs models on your behalf. Cursor used Fireworks AI for this and issued it an access key."},{"term":"Marketplace cache","def":"The copy of an extension served to users. Poisoning the cache serves a backdoored build without touching the original publisher."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Vendor keys are the way in","text":"One key at one inference vendor reached source, support data and users.","abstain":false},{"id":"b","label":"One vendor's lapse, contained","text":"Under 0.01% of customers had credentials exposed. Cursor cut Fireworks off.","abstain":false},{"id":"c","label":"Agent traces are the prize","text":"The bucket held 2025 traces with unredacted credentials inside them.","abstain":false},{"id":"d","label":"Ordinary supply chain crime","text":"Marketplace caches get poisoned. The AI part is incidental.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"c","direction":"strengthens","observable_event":"A second AI coding product discloses credentials exposed through stored agent traces, or Cursor, Anthropic, OpenAI or Google ships redaction of secrets at write time, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"strengthens","observable_event":"Fireworks AI or another inference vendor is named as the origin of a further customer breach, or a regulator opens an inquiry into AI subprocessor access, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"The investigations close with no further affected customers named and no additional vendor key implicated, and Fireworks publishes a final report, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"}]}