{
  "type": "NewsArticle",
  "url": "https://sonomos.ai/newsroom/fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking/",
  "slug": "fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking",
  "title": "Fireworks AI breach widens: Notion AI chat transcripts, billing data and model weights taken",
  "summary": "Fireworks AI says an attacker who reached its development environment through a cloud provider took billing data and, for some accounts, datasets and incomplete model weights. Notion says Notion AI chat transcripts were included.",
  "section": "Cybersecurity",
  "category": "cybersecurity",
  "format": "analysis",
  "tags": [
    "AI security",
    "Data breach",
    "Third-party risk",
    "Prompt leakage",
    "LLMjacking"
  ],
  "published_at": "2026-11-15T14:11:03.000Z",
  "updated_at": "2026-11-15T14:11:03.000Z",
  "word_count": 855,
  "publisher": {
    "name": "Sonomos Newsroom",
    "url": "https://sonomos.ai/newsroom/",
    "standards": "https://sonomos.ai/newsroom/standards/",
    "organization": {
      "name": "Sonomos",
      "url": "https://sonomos.ai/"
    }
  },
  "follows_up": null,
  "followed_by": [
    {
      "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",
      "url": "https://sonomos.ai/newsroom/cursor-discloses-fireworks-key-coding-agent-traces-malicious-extension/",
      "published_at": "2026-11-16T14:17:27.000Z"
    }
  ],
  "sources": [
    {
      "number": 1,
      "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": "2026-11-15"
    },
    {
      "number": 2,
      "title": "Security incident involving a third-party AI vendor",
      "publisher": "Notion",
      "url": "https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor",
      "published": null,
      "accessed": "2026-11-15"
    },
    {
      "number": 3,
      "title": "Cyberattack on Fireworks AI allowed hackers to access Notion customers' chat transcripts",
      "publisher": "TechCrunch",
      "url": "https://techcrunch.com/2026/11/12/cyberattack-on-fireworks-ai-allowed-hackers-to-access-notion-customers-chat-transcripts/",
      "published": "2026-11-12",
      "accessed": "2026-11-15"
    },
    {
      "number": 4,
      "title": "The Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model Resources",
      "publisher": "Google Cloud (Mandiant and Google Threat Intelligence Group)",
      "url": "https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models",
      "published": "2026-11-02",
      "accessed": "2026-11-15"
    }
  ],
  "verification": {
    "method": "automated-source-check",
    "checked_at": "2026-11-15T14:11:03.000Z",
    "description": "https://sonomos.ai/newsroom/standards/",
    "claims": [
      {
        "claim": "Fireworks says an attacker reached its development machines on October 24 after a third-party cloud provider hosting its R&D machines had an incident on October 23.",
        "source": 1,
        "source_url": "https://fireworks.ai/security/bulletins/october-2026-security-incident-update"
      },
      {
        "claim": "Fireworks lists resource metadata, billing information, and for a limited set of accounts secrets, datasets and incomplete model weights as exposed; card numbers were not affected.",
        "source": 1,
        "source_url": "https://fireworks.ai/security/bulletins/october-2026-security-incident-update"
      },
      {
        "claim": "Notion says Notion AI chat transcripts from its Insights feature were affected, for fewer than 1% of workspaces, and that it disconnected from Fireworks on November 3.",
        "source": 2,
        "source_url": "https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor"
      },
      {
        "claim": "Google's Mandiant and GTIG describe LLMjacking as hijacking enterprise LLM access to run inference at the victim's expense, with some victims billed over $100,000 in a day.",
        "source": 4,
        "source_url": "https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models"
      }
    ]
  },
  "corrections": [],
  "body_markdown": "Fireworks AI, a platform developers use to run open AI models, said in a security bulletin last updated on November 9 that the incident it first disclosed in October has widened: an attacker who reached its development environment through a third-party cloud provider took account and billing information and, for a limited set of accounts, secrets, datasets and incomplete model weights. Notion said the exposed data also included chat transcripts from its Notion AI \"Insights\" feature, affecting fewer than 1% of Notion workspaces.\n\n## What did Fireworks say was taken?\n\nThe [Fireworks bulletin](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) gives a timeline. The company first reported unauthorized use of internal credentials in early October, when it said only secrets used in evaluation jobs were known to be accessed. The update says a third-party cloud provider hosting Fireworks research and development machines had an incident on October 23, and that on October 24 an unauthorized party reached Fireworks development machines and ran a workload that caused additional customer data impact. Fireworks says it detected this the same day and found no further evidence of unauthorized access to the platform customers use by October 25.\n\n| Category | What Fireworks lists |\n|---|---|\n| Resource metadata | Identifiers, display names, descriptions, creation times |\n| Billing information | Billing and shipping addresses, invoice and payment history, account balance, credit usage |\n| Customer content, limited set of accounts | Secrets, datasets, and a subset of incomplete model weights |\n| Not affected | Payment card numbers, which Fireworks says it does not store |\n\nFireworks says it found no impact on Inference, Training SDK or Nexus, which it describes as zero data retention by default, and that this reflects its investigation to date. It has not published how many customers were affected, saying only that impacted customers are being notified. It says it notified law enforcement and engaged Palo Alto Networks Unit 42 for forensic analysis.\n\n## What did Notion say was exposed?\n\n[Notion's incident page](https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor) says Fireworks told it of an incident affecting fewer than 1% of Notion workspaces. The exposed data, it says, is the full text of users' Notion AI chat messages, which may include names, email addresses, API tokens or other information users typed, plus the text of the AI's output, which may include fragments of referenced pages. For most affected transcripts, Notion says, the metadata linking a thread to a user or workspace was not present. AI Meeting Notes were not affected unless their content was typed into a chat or referenced in AI output.\n\nNotion says Insights used Fireworks for classification jobs, and that customers in its Learning and Early Access Program had volunteered data to experiment with training task-specific models there. It says Fireworks had contractually committed to zero data retention, but that Notion later learned data was retained from the endpoints it used. [TechCrunch reported](https://techcrunch.com/2026/11/12/cyberattack-on-fireworks-ai-allowed-hackers-to-access-notion-customers-chat-transcripts/) that Notion disconnected its systems from Fireworks on November 3; Notion's page says all of its data was deleted from Fireworks' systems on November 4, the Insights feature was disabled, and it no longer serves open-weight models through Fireworks. Notion also says it is auditing critical vendors for data minimization and zero-data-retention practices.\n\n## Who is behind the attack, and what is LLMjacking?\n\nFireworks says evidence links the incident to a financially motivated group that sells tokens obtained through unauthorized access to illicit routing gateways, activity the bulletin calls LLMJacking. Google's Mandiant and Google Threat Intelligence Group published a [report on the practice](https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models) on November 2. They describe attackers hijacking enterprise access to language models and cloud AI services and running inference at the victim's expense, with single-day charges above $100,000 and cumulative charges reaching $1 million at some organizations before access was revoked. The report names exposed or stolen cloud credentials as the dominant way in and describes \"LLM mixers\" that pool stolen access and resell it. The Google report does not mention Fireworks.\n\n## Why does the vendor chain matter here?\n\nThe incident passed through three layers: a cloud provider, an inference vendor, and an application that sent users' prompts to it. Notion's customers were affected through the second and third, and the Notion page says the retention the vendor had promised was not what occurred. That is a different failure from a flaw in the model itself. It also follows other recent AI infrastructure disclosures, including the [NEAR AI SGLang breach](https://sonomos.ai/newsroom/near-ai-sglang-breach-confidential-computing-account-data-exposed/), where the material at risk was likewise data and credentials held around the model, and the [Chainalysis breach](https://sonomos.ai/newsroom/chainalysis-breach-offensive-agentic-techniques-user-data-exposed/).\n\nFor readers weighing what typed prompts can contain, Notion's own list (names, emails and API tokens) is an example of [prompt leakage](https://sonomos.ai/glossary/prompt-leakage/) arising downstream of the user, at a vendor. Sonomos has written about how [AI agents change the data privacy picture](https://sonomos.ai/blog/ai-agents-data-privacy-2026/).\n\n## What is unresolved?\n\nFireworks says its investigation is ongoing and that it is in active monitoring. It has not said how many accounts had datasets or model weights taken, which customers they are, or how long the attacker had access to the development environment. TechCrunch names Harvey and Cursor as Fireworks customers but the sources do not say either was affected. Notion says its investigation is continuing and that it will share updates. Neither company's page mentions regulatory notifications.",
  "links": {
    "html": "https://sonomos.ai/newsroom/fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking/",
    "markdown": "https://sonomos.ai/newsroom/fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking/index.md",
    "json": "https://sonomos.ai/newsroom/fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking/index.json",
    "image": "https://sonomos.ai/og/newsroom/fireworks-ai-breach-notion-ai-chat-transcripts-llmjacking.png"
  }
}
