Get a demo — 30 minutes →
← Back to blog
Anomity robot illustrating The Fireworks AI Breach - How Stored Agent Traces Led to Cursor's Malicious Claude Code Extension
Research

The Fireworks AI Breach - How Stored Agent Traces Led to Cursor's Malicious Claude Code Extension

TL;DR
  • On November 13, 2026 Cursor said the malicious copy of Anthropic's Claude Code extension served from its marketplace on October 29 began at Fireworks AI, a model training and inference provider. An access key Cursor had issued to Fireworks was exposed in the Fireworks breach, and the attacker used it to copy a storage bucket of coding agent traces.
  • A small number of those traces held credentials that had not been redacted. Starting from them, the attacker reached SaaS tools Cursor uses, then used a stolen credential to place tampered builds of the extension on the cache that serves Cursor's extension marketplace. Cursor says at most 5,000 users may have received the build during about 11 hours.
  • Cursor is not the only downstream victim. Notion says Notion AI chat transcripts from its Insights feature were in a Fireworks dataset, affecting under 1% of workspaces, and it is revoking API keys found in those chats. TechCrunch reports that Uber is notifying drivers whose license data was taken in the same breach.
  • Fireworks ties the intrusion to a financially motivated group that sells stolen access to large language models (LLMjacking). It says secrets, datasets and partial model weights were reached for a limited set of accounts, but it has seen no impact to its zero-retention inference.
  • The lesson for security teams: whatever people paste into an agent or AI chat can outlive the session, in training sets, evaluation jobs and analytics stores at vendors you may not know about. Treat credentials typed into AI tools as exposed, and know which AI extensions and versions sit on each machine.

The short version: one of the most consequential AI supply-chain incidents of this autumn did not start with a poisoned package or a prompt injection. It started with old coding agent sessions sitting in a storage bucket at an AI vendor, and with the passwords and API keys that people had typed into those sessions. On November 13, 2026 Cursor published the results of its investigation into the trojanized Claude Code extension served from its marketplace on October 29. The answer to the question our advisory left open, how the attacker got the access to place that build, runs back through Fireworks AI.

Fireworks hosts open models for inference and fine-tuning, and many AI products send it data. Cursor, Notion and Uber have now all described data that reached the attacker through it. Each disclosure on its own is a vendor incident. Read together, they show a pattern security teams should plan for: agent traces and chat transcripts are a store of secrets, and copies of them travel to third parties for training, evaluation and analytics.

What to do now

  • If you use Cursor, check whether your team admins got a notice from Cursor. Cursor says it emailed admins a list of specific keys to rotate where credentials were exposed. If anyone installed or updated to anthropic.claude-code 2.1.312 or 2.1.313 from Cursor's marketplace during the exposure window, follow Cursor's security notice. Cursor warns that uninstalling the extension alone may not be enough.
  • If you use Notion AI, check whether your workspace owner received a notice. Notion recommends treating the affected chats as exposed and rotating any passwords, API keys or other credentials pasted into Notion AI during the incident window.
  • If you are a Fireworks customer, act on the account-specific notice. Fireworks says removing a stored secret does not invalidate it, so third-party credentials must be rotated with the providers that issued them.
  • Assume pasted credentials are exposed. Any key a developer pastes into a coding agent or an AI chat can end up in a transcript, a trace or a training set. Rotate on paste, not on breach.
  • Ask your AI vendors where your data goes after the request. "Zero data retention" usually covers live inference. Training sets, evaluation jobs, batch jobs and analytics features can store data, and the subprocessors behind them may hold copies.
  • Keep an inventory of AI extensions by version. When a marketplace serves a bad build, the first question is which machines installed which version, and when.

How the chain reached Cursor's marketplace

Cursor's update describes the attack in stages. Below is our summary of that account, with dates from Cursor and Fireworks.

StageWhat happenedSource
Vendor breachUnauthorized use of internal Fireworks credentials, identified on October 6. A third-party cloud provider that hosted Fireworks R&D machines had its own incident on October 23, and the attacker used it to reach Fireworks development machines on October 24.Fireworks bulletins
Shared bucketAn access key Cursor had issued to Fireworks was exposed in that breach. The attacker used it to copy a read-only bucket Cursor used to share data with Fireworks.Cursor
Agent tracesThe bucket held limited coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which Cursor was permitted to use for training. A small number held credentials that had not been redacted.Cursor
SaaS pivotWith those credentials the attacker reached select SaaS tools used by Cursor: archived non-production repositories, some support data and email, usage metrics and contact details, summaries of some agent conversations, and credentials users had sent to Cursor support.Cursor
MarketplaceOn October 29, a stolen credential from that data let the attacker place tampered copies of the Claude Code extension on the cache serving Cursor's VS Code extension marketplace. Open VSX, the upstream source, never had the malware.Cursor
ContainmentCursor says the build was live for about 11 hours and reached at most 5,000 users. The last attacker activity it saw was on November 1, and it was blocked.Cursor

Cursor puts the scale this way: fewer than 0.01% of customers had API credentials exposed in the traces, and traces were exposed for fewer than 0.05% of users who were not in Privacy Mode. It says users with Privacy Mode on, customer code repositories, its production systems, the Cursor app and its official updates, and every other extension and plugin in its marketplace were not affected, and that no release signing keys were exposed. Cursor has moved all inference off Fireworks, rotated every exposed credential and, as a precaution, all credentials in use across its internal systems, restricted where service-provider credentials can be used from, and extended secret scanning beyond code repositories to other internal data sources.

That last fix is the telling one. Secret scanning usually points at code. Here the secrets that opened the door sat in agent session data, which is exactly the kind of data that gets copied out for model work.

Notion and Uber: the same vendor, different data

Notion says it used Fireworks for classification jobs behind Insights, a feature that helps customers understand their own AI usage, and as an inference provider for open-weight models. Notion's timeline: Fireworks first told it on November 2 that a small dataset of synthetic test data had been accessed. Notion disconnected from Fireworks by the morning of November 3 and deleted its data there on November 4. On November 6 Fireworks said customer data may have been accessed, and on November 8 it confirmed the final dataset, which held Notion AI chat transcripts used for Insights.

Notion says the incident affects fewer than 1% of workspaces, and that for most of them the transcripts are unattributable: not labeled with a user or workspace, and without tool call results. The text people typed is still there, though, which is why Notion revoked Notion API keys found in affected messages on November 11 and, on November 12, reported third-party API keys found in those messages to the vendors that issued them. Insights is disabled while the investigation continues, and Notion no longer serves open-weight models through Fireworks.

Uber told regulators in a notice dated November 10 that a service provider informed it on October 9 that an unauthorized party had used the provider's compromised internal credentials to access Uber's account with that provider. Uber's notice does not name the provider. An Uber spokesperson told TechCrunch that it was Fireworks, and TechCrunch counted driver's license data for 2,567 people in Massachusetts, 3,430 in South Carolina and 1,873 in Rhode Island in state filings, with Illinois' count still pending.

What Fireworks says happened

Fireworks' first bulletin, from October 9, said environment variables used in evaluation jobs, including API keys and tokens stored as secrets, were captured in a limited number of customer accounts. Its November 9 update widened that: account and billing information was affected, and for a limited set of accounts so were secrets, datasets and incomplete model weights. Fireworks says it is working with Palo Alto Networks Unit 42, has notified law enforcement, and has paused self-serve dataset uploads and downloads for batch inference and dataset-based training while it investigates.

Fireworks says the evidence links the intrusion to a financially motivated group that sells access to large language models through illicit routing gateways, the model we describe in our LLMjacking advisory. It says it has seen no impact to inference, its Training SDK or Nexus, which are zero data retention by default. The gap between that statement and Notion's experience is the point: zero retention described the live inference path, while the data that leaked sat in stored datasets used for other jobs.

Why agent traces are a secrets problem

A coding agent session is a full record of work: the files read, the commands run, the errors pasted back, and often the token someone dropped in to get a failing call working. An AI chat transcript is similar. People treat both as a conversation, but to everyone who later stores, labels or trains on that record, it is data, and it can hold live credentials. We saw the local version of the same problem in Apiiro's research on Claude Code saving secrets into permission files. The Fireworks chain is the remote version: the secrets left the laptop inside the session itself.

Two other details matter for defenders. First, the attacker did not need to compromise Anthropic or Open VSX to ship a trusted-looking extension. A credential to one editor vendor's marketplace cache was enough, which is why approving an extension by name is weaker than knowing which build each machine runs. Second, Privacy Mode did what Cursor says it does, keeping those users' sessions out of the shared data. As we noted in our look at Cursor Privacy Mode and enterprise controls, it is a data-handling guarantee worth enforcing on every seat, and it governs where data goes, not what is on the endpoint.

How Anomity surfaces and governs it

See the extensions and agents. The Anomity Endpoint Sensor records every installed Cursor and VS Code extension with its publisher, ID and version on each managed Windows, macOS and Linux machine, alongside the AI agents, MCP servers, plugins, skills, hooks and CLIs around it. Finding every device that took anthropic.claude-code 2.1.312 or 2.1.313 is a query, and the 90-day audit trail shows when each one moved onto and off a build.

See the secrets. The sensor reports which credentials and AI provider keys sit on each machine by type and location, with values redacted on the endpoint, so a rotation list starts from a record. The Anomity Browser Sensor covers AI chats in the browser: which AI services are used, whether the signed-in account is corporate or personal, and which credentials were pasted or typed into them, with the option to redact or block on the way in. Cloud discovery adds the Google Workspace and GitHub OAuth grants that AI apps hold.

Enforce before the call runs. On agents that expose a hook, such as Claude Code PreToolUse, Anomity's managed hooks can allow, deny or log a tool call before it runs, for example to stop an agent from reading credential files it has no reason to touch. Findings route to SIEM, Slack, email or Jira. Anomity is SOC 2 Type II and complements EDR, DLP, network and GRC tooling; it does not replace vendor risk reviews. To see which AI extensions, agents and secrets are on your fleet today, book a 30-minute demo.

Frequently asked questions

What caused Cursor's malicious Claude Code extension incident?

Cursor said on November 13, 2026 that it began with an access key Cursor had issued to Fireworks AI. The key was exposed in a breach at Fireworks, and the attacker used it to copy a storage bucket of coding agent traces. Some traces held unredacted credentials, which led to SaaS tools Cursor uses, and a stolen credential from that data let the attacker place tampered copies of Anthropic's Claude Code extension on the cache serving Cursor's extension marketplace on October 29.

Was Anthropic or Open VSX compromised?

No. Cursor says Anthropic did not publish the copy, it was not a vulnerability in Anthropic's product, and the marketplace source, Open VSX, never had the malware. The tampered build was served from the cache behind Cursor's own marketplace.

How many people were affected?

Cursor says at most 5,000 users who requested specific cached versions of the extension on October 29 may have received the malicious build, which was live for about 11 hours. It says fewer than 0.01% of customers had API credentials exposed in the traces. Notion says fewer than 1% of its workspaces were affected. Uber has not given a total, and TechCrunch counted several thousand drivers in state filings.

Which extension versions should I check?

Cursor's update asks anyone who installed or updated to anthropic.claude-code 2.1.312 or 2.1.313 from the Cursor VS Code extension marketplace during the exposure window to follow the steps in its security notice, and warns that uninstalling the extension alone may not be enough.

Who is behind the Fireworks AI breach?

Fireworks says the evidence links it to a financially motivated cybercriminal group that sells unauthorized access to large language models to illicit routing gateways, known as LLMjacking. No group has been publicly named or claimed the attack.

Did Cursor Privacy Mode protect users?

According to Cursor, yes. The exposed traces came from sessions not in Privacy Mode, which Cursor was permitted to use for model training, and Cursor lists prompts and code from Privacy Mode users as not affected.

What should security teams change after this incident?

Treat credentials pasted into coding agents and AI chats as exposed and rotate them, ask AI vendors which features store data outside zero data retention and which subprocessors receive it, enforce privacy settings such as Cursor Privacy Mode on every seat, and keep a per-machine inventory of AI extensions and versions so a bad marketplace build can be scoped quickly.

How does Anomity help with this?

Anomity's Endpoint Sensor records every installed Cursor and VS Code extension with its version on each managed machine, along with AI agents, MCP servers, skills, hooks and the secrets present, with values redacted on the endpoint. The Browser Sensor shows which AI services employees use and which credentials were pasted into them. Hooks on agents such as Claude Code can allow, deny or log tool calls before they run, and a 90-day audit trail keeps the record.

Ask AI about Anomity
ChatGPT Claude Perplexity Google AI Grok