{
  "type": "NewsArticle",
  "url": "https://sonomos.ai/newsroom/cursor-discloses-fireworks-key-coding-agent-traces-malicious-extension/",
  "slug": "cursor-discloses-fireworks-key-coding-agent-traces-malicious-extension",
  "title": "Cursor traces breach to exposed Fireworks key: agent traces with live credentials, then a malicious extension",
  "summary": "Cursor says an access key it issued to AI vendor Fireworks was exposed, letting an attacker copy coding-agent traces that held unredacted credentials and later plant a malicious extension build that reached at most 5,000 users.",
  "section": "Cybersecurity",
  "category": "cybersecurity",
  "format": "analysis",
  "tags": [
    "AI security",
    "Data breach",
    "Third-party risk",
    "Prompt leakage"
  ],
  "published_at": "2026-11-16T14:17:27.000Z",
  "updated_at": "2026-11-16T14:17:27.000Z",
  "word_count": 811,
  "publisher": {
    "name": "Sonomos Newsroom",
    "url": "https://sonomos.ai/newsroom/",
    "standards": "https://sonomos.ai/newsroom/standards/",
    "organization": {
      "name": "Sonomos",
      "url": "https://sonomos.ai/"
    }
  },
  "follows_up": {
    "slug": "fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking",
    "title": "Fireworks AI breach widens: Notion AI chat transcripts, billing data and model weights taken",
    "url": "https://sonomos.ai/newsroom/fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking/",
    "published_at": "2026-11-15T14:11:03.000Z"
  },
  "followed_by": [],
  "sources": [
    {
      "number": 1,
      "title": "An update on our recent security incident",
      "publisher": "Cursor",
      "url": "https://cursor.com/blog/security-incident-update",
      "published": "2026-11-13",
      "accessed": "2026-11-16"
    },
    {
      "number": 2,
      "title": "Malicious Anthropic.claude-code extension in the Cursor marketplace: security notice",
      "publisher": "Cursor",
      "url": "https://cursor.com/blog/claude-code-extension-security-notice",
      "published": "2026-10-29",
      "accessed": "2026-11-16"
    },
    {
      "number": 3,
      "title": "Security bulletin: October 2026 security incident update",
      "publisher": "Fireworks AI",
      "url": "https://fireworks.ai/security/bulletins/october-2026-security-incident-update",
      "published": "2026-11-09",
      "accessed": "2026-11-16"
    }
  ],
  "verification": {
    "method": "automated-source-check",
    "checked_at": "2026-11-16T14:17:27.000Z",
    "description": "https://sonomos.ai/newsroom/standards/",
    "claims": [
      {
        "claim": "Cursor says an attacker used an access key it had issued to Fireworks, which Fireworks had exposed, to copy data from a read-only storage bucket shared with Fireworks.",
        "source": 1,
        "source_url": "https://cursor.com/blog/security-incident-update"
      },
      {
        "claim": "Cursor says at most 5,000 users who requested specific cached versions on October 29 could have received the malicious extension build, which was available for about 11 hours.",
        "source": 1,
        "source_url": "https://cursor.com/blog/security-incident-update"
      },
      {
        "claim": "Cursor says less than 0.01% of customers had API credentials exposed and fewer than 0.05% of users had agent traces exposed.",
        "source": 1,
        "source_url": "https://cursor.com/blog/security-incident-update"
      },
      {
        "claim": "Fireworks says a third-party cloud provider hosting its R&D machines had an incident on October 23 and that development machines were reached on October 24.",
        "source": 3,
        "source_url": "https://fireworks.ai/security/bulletins/october-2026-security-incident-update"
      }
    ]
  },
  "corrections": [],
  "body_markdown": "Cursor, the AI coding-editor maker, said on November 13 that an attacker who used an access key it had issued to AI inference vendor Fireworks copied coding-agent traces from a shared storage bucket, found unredacted credentials in them, and used those to reach internal tools and, on October 29, place malicious copies of the Claude Code extension in its marketplace cache. This is a development in the Fireworks incident [we covered on November 15](https://sonomos.ai/newsroom/fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking/): it is the first account from one of Fireworks' customers of how the stolen access was used downstream.\n\n## What did Cursor say happened?\n\nIn its [November 13 update](https://cursor.com/blog/security-incident-update), Cursor says the attacker used a key it had issued to Fireworks, which Fireworks had exposed, to copy data from a read-only storage bucket the two companies used to share data. Cursor says the bucket held limited coding-agent traces, mostly from 2025, from sessions not in Privacy Mode, which Cursor was permitted to use for model training. Some traces contained credentials that had not been redacted.\n\nCursor says the attacker used those credentials to reach select software-as-a-service tools. It lists what was accessed: archived non-production source repositories, some with business contact information; limited customer support data and email conversations; product usage metrics and contact details; summaries of some agent conversations; and credentials users had sent to Cursor support.\n\n## How many people were affected?\n\nCursor gives several percentages and counts that measure different things, so they should not be added together.\n\n| Measure | Cursor's figure |\n|---|---|\n| Customers with API credentials exposed | Less than 0.01% |\n| Users with agent traces exposed (outside Privacy Mode) | Fewer than 0.05% |\n| Users who could have received the malicious extension build | At most 5,000 |\n| Time the malicious build was available | About 11 hours |\n\nCursor says affected customers were contacted directly, and that most customers were not affected.\n\n## What did the malicious extension do?\n\nCursor's [October 29 security notice](https://cursor.com/blog/claude-code-extension-security-notice) says versions 2.1.312 and 2.1.313 of the Anthropic.claude-code extension, for macOS and Windows, were malicious and were available from about 2:27 a.m. to 1:35 p.m. Pacific that day. The notice says they collected host and user details (hostname, username, operating system, home path, environment variable names, Git email and public IP address) and tried to add an attacker's SSH key, enable remote login and open a reverse tunnel, and to make that persistent. Cursor says the SSH step would usually fail without administrator rights.\n\nThe November update says Anthropic did not publish the copy and that the issue was not a vulnerability in Anthropic's product, and that the upstream Open VSX registry never held the malware. The \"at most 5,000\" figure is an upper bound on users who requested those cached versions, not a count of confirmed infections. Cursor's notice says uninstalling the extension alone may not be enough, because a key or persistence mechanism could remain, and it published indicators of compromise and a remediation script.\n\n## What does this show about how agent data travels?\n\nThe [Fireworks bulletin](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) says a third-party cloud provider hosting its research machines had an incident on October 23, and that an unauthorized party reached its development machines on October 24. It does not say how the key Cursor issued came to be exposed, and Cursor's account does not either; the two companies' descriptions have not been reconciled in public.\n\nWhat Cursor describes is a chain: a vendor held a key, the key opened a bucket, the bucket held agent traces, and the traces held live secrets. Agent traces record what a coding agent read, ran and was shown, so they can hold material a developer never meant to share, the same pattern as [prompt leakage](https://sonomos.ai/glossary/prompt-leakage/) but through logs rather than a typed prompt. Cursor says Privacy Mode users' prompts and code were not affected, which is a data-handling setting doing what it was meant to. Sonomos has written about [how AI agents change the data privacy picture](https://sonomos.ai/blog/ai-agents-data-privacy-2026/).\n\n## What did Cursor change?\n\nCursor says it moved all inference traffic off Fireworks, deactivated the key and cut Fireworks' access to shared datasets. It says it rotated every exposed credential it found and, as a precaution, all credentials in internal use; validated every other marketplace extension; rebuilt the marketplace; added secret scanning across code repositories and other internal data sources; redeployed production infrastructure; and engaged external forensic experts. It says the last attacker activity it saw was on November 1 and was blocked, and that it regards the incident as contained.\n\n## What is unresolved?\n\nCursor says it will update its post and contact customers if it finds more. Fireworks says its investigation is ongoing. Neither company says how many people ran the malicious build, whether any regulator has been notified, or how the Fireworks exposure of Cursor's key occurred. Cursor's account is its own investigation's finding and has not been independently confirmed in the sources reviewed here.",
  "links": {
    "html": "https://sonomos.ai/newsroom/cursor-discloses-fireworks-key-coding-agent-traces-malicious-extension/",
    "markdown": "https://sonomos.ai/newsroom/cursor-discloses-fireworks-key-coding-agent-traces-malicious-extension/index.md",
    "json": "https://sonomos.ai/newsroom/cursor-discloses-fireworks-key-coding-agent-traces-malicious-extension/index.json",
    "image": "https://sonomos.ai/og/newsroom/cursor-discloses-fireworks-key-coding-agent-traces-malicious-extension.png"
  }
}
