Fireworks AI Data Breach Widens to Model Weights, Billing Data

Fireworks AI widened its breach disclosure to billing data, secrets, and incomplete model weights; Notion confirms exposed AI Insights chat transcripts.

The Fireworks AI logo on a dark red gradient background, the company's official brand image.
Image: Fireworks AI.

Fireworks AI has told customers that a breach it first disclosed in October 2026 was bigger than the company first said. In an updated security bulletin published November 9, the AI inference provider said attackers who broke in through a compromised third-party cloud provider also reached billing and account data, resource metadata, and — for a limited number of accounts — secrets, incomplete model weights, and datasets. Separately, Notion has confirmed that chat transcripts from its Fireworks-powered "AI Insights" feature were exposed, affecting fewer than 1% of Notion workspaces.

Key facts
  • Fireworks AI's October 9 bulletin first described stolen credentials reaching secrets used in evaluation jobs.
  • A November 9 update says a separate intrusion into R&D machines, via a breached third-party cloud provider, led to wider exposure: billing/account data, resource metadata, and — for a limited set of accounts — secrets, incomplete model weights, and datasets.
  • Fireworks says payment card numbers were not affected because it does not store them, and Inference, the Training SDK, and Nexus (all zero-data-retention by default) show no detected impact.
  • Notion separately disclosed that AI Insights chat transcripts sent to Fireworks were exposed, affecting fewer than 1% of Notion workspaces; Notion says this was not a breach of its own systems.
  • Fireworks has not detailed exactly what "incomplete model weights" means, which specific datasets were involved, or whether any regulated or sensitive data was in scope.

What happened

Fireworks AI, a platform that hosts and serves open-weight AI models for companies including Notion and reportedly Cursor, first disclosed a security incident on October 9, 2026. That initial bulletin said a threat actor had used stolen internal credentials to reach secrets Fireworks stores for customers, such as those used in evaluation jobs.

The picture changed after a second intrusion. According to Fireworks' own updated security bulletin, a third-party cloud provider that hosts Fireworks' research-and-development machines was itself breached on October 23. The attacker used that foothold to reach Fireworks development machines on October 24 and ran a workload that caused additional customer-data impact — which Fireworks says it detected the same day. By October 25, Fireworks reported no further evidence of successful unauthorized access to its customer-facing platform.

Fireworks says it has worked with incident-response firm Palo Alto Networks' Unit 42, notified law enforcement, and links the activity to a financially motivated group that resells stolen LLM access through underground marketplaces — a technique researchers call "LLMJacking." The company's November 9 update is, by its own account, still not final: "the investigation is ongoing and may reveal more information."

Timeline of the Fireworks AI breach

Piecing together Fireworks' own disclosures, the sequence runs:

DateEvent
Oct 9, 2026Fireworks publishes its first bulletin, citing stolen credentials and exposure of secrets such as those used in evaluation jobs.
Oct 23, 2026A third-party cloud provider hosting Fireworks' R&D machines is breached, per Fireworks' account.
Oct 24, 2026The attacker uses that access to reach Fireworks development machines and runs a workload that widens customer-data impact; Fireworks detects it the same day.
Oct 25, 2026Fireworks reports no further evidence of successful unauthorized access to the customer-facing platform.
Nov 2, 2026Google's Threat Intelligence Group and Mandiant publish a report on "LLMJacking" — the broader trend Fireworks says this attack fits.
Nov 8–9, 2026Fireworks confirms the wider dataset to customers, including Notion; Fireworks and Notion each publish updated disclosures.

What data was exposed

Fireworks' updated bulletin breaks the exposure into several categories. Not every category applies to every customer — the company is explicit that model weights, datasets, and secrets were accessed only "for a limited set of accounts," while billing and account information is described more broadly.

Data categoryStatus per FireworksDetail
Payment card numbersNot affectedFireworks says it does not store card numbers, so they were not impacted.
Billing & account dataAffectedBilling and shipping addresses, invoice and payment history, account balance, and credit usage.
Resource metadataPotentially accessibleThe affected internal application could reach identifiers, display names, descriptions, and creation times; Fireworks does not confirm this metadata was actually accessed.
SecretsAffected for a limited set of accountsExpands on the original bulletin, which described secrets used in evaluation jobs.
Model weightsAffected for a limited set of accountsA subset of incomplete model weights was accessed; Fireworks has not specified what "incomplete" means in practice.
DatasetsAffected for a limited set of accountsFireworks has not named which datasets or how many customers.
Inference, Training SDK, NexusNo detected impactAll three run with zero data retention (ZDR) by default, per Fireworks.

Fireworks has disabled self-serve dataset uploads and downloads for Batch Inference and dataset-based training while the investigation continues. The Training SDK remains available under zero data retention, and the company says lab-based training now runs in an isolated environment.

The Notion angle: AI Insights chat transcripts

Fireworks counts Notion among the customers affected, and Notion has published its own account of what that meant for its users. Notion's security advisory, also dated November 9, lays out a separate timeline on its side:

  • On the evening of November 2, Fireworks told Notion that malicious actors had accessed a small set of data. Notion's investigation found it was synthetic testing data with no customer data involved.
  • By the morning of November 3, Notion says it had, "out of an abundance of caution," disconnected its systems from Fireworks, deleted all data in the Fireworks system, and stopped communicating with Fireworks.
  • On November 6, Fireworks told Notion that a second, potentially customer-impacting dataset may have been accessed. Notion opened a new internal investigation.
  • Fireworks confirmed that dataset to Notion on November 8, and Notion published its advisory the next day.

The data at issue is chat transcripts from Notion AI's "Insights" feature, which used Fireworks-hosted models. Notion says the incident affected fewer than 1% of its workspaces, and that it has already contacted customers it identified as affected — a workspace that hasn't heard from Notion, the company says, has not been identified as impacted. For most of the exposed transcripts, Notion says they were not labeled with an author or workspace and did not include tool-call results (such as the contents of pages Insights might have referenced), which limits how easily any single conversation could be tied back to a specific user or company.

Notion is explicit that it considers this a vendor incident, not a breach of Notion's own systems: "this was not a breach of Notion's systems." In response, Notion disabled the Insights feature, said it has stopped serving open-weight models through Fireworks, and emailed affected workspace owners with instructions for reviewing the exposed data. Notion also advises any user who pasted sensitive information — passwords, API keys, or other credentials — into Notion AI during the incident window to rotate those credentials, similar to the standard advice after other third-party breaches, such as the Apify breach traced to stolen credentials earlier this year.

What's still unclear

Fireworks' disclosures are more detailed than most vendor breach bulletins, but several important questions remain open as of this writing:

  • What "incomplete model weights" actually means. Fireworks has not said whether this refers to partial checkpoints, truncated files, weights without full training metadata, or something else — nor how usable any accessed weights would be to a third party.
  • Which datasets, and whose. Fireworks has not named the datasets involved or said how many customer accounts fall into the "limited set" affected by the weights, dataset, and secrets exposure.
  • Whether sensitive or regulated data was involved. Neither Fireworks' bulletin nor Notion's advisory states whether any exposed data falls under frameworks like GDPR, HIPAA, or similar regimes, or whether regulators have been notified beyond the law-enforcement contact Fireworks mentions.
  • Full customer scope. Notion and (per earlier reporting) Cursor have both been named as affected Fireworks customers, but Fireworks' own bulletin does not publish a customer list, so the total number and identity of affected organizations beyond what those companies choose to disclose is not public.

Until Fireworks or affected customers provide more specifics, readers should treat "incomplete model weights and datasets" as a real but still loosely defined category of exposure — serious enough that Fireworks disabled related self-serve features, but not yet detailed enough to assess practical risk.

Fireworks' response and remediation

Per its bulletin, Fireworks has taken a series of remediation steps since detecting the broader intrusion:

  • Revoked and replaced affected credentials and service accounts, and removed unnecessary elevated privileges.
  • Restricted unauthorized network paths and isolated affected systems.
  • Reviewed and tightened access controls on data storage and analytics systems.
  • Strengthened logging and monitoring, and says it is now monitoring systems 24/7.
  • Disabled self-serve dataset uploads and downloads for Batch Inference and dataset-based training during the investigation.
  • Continued working with Unit 42 and notified law enforcement.

Fireworks says it is notifying affected customers directly with recommended actions and points people with questions to [email protected] or its Trust Center.

Why this keeps happening: "LLMJacking"

Fireworks ties the attack to a trend researchers have taken to calling "LLMJacking": financially motivated actors who steal cloud or API credentials specifically to resell unauthorized access to AI models through underground gateways, rather than to mine cryptocurrency or exfiltrate data for its own sake. A report published November 2 by Google's Threat Intelligence Group and Mandiant — which Fireworks' bulletin cites — describes this activity across victims in financial services, retail, technology, and transportation, including organizations that say they had never knowingly enabled LLM services. The report also notes that this kind of abuse often slips past usage-based anomaly alerts built into conventional EDR and SIEM tooling, since a sudden spike in AI API calls doesn't necessarily look like a conventional breach pattern to systems tuned for other threats.

That broader pattern is a useful frame for the Fireworks incident: the initial entry point was stolen credentials, not a flaw in Fireworks' AI models themselves, and the attacker's apparent goal — reaching compute and model access to resell — is different from, say, a straightforward credential-stuffing attack aimed at account takeover. It doesn't make the resulting exposure of billing data, secrets, and model assets any less real for affected customers, but it does help explain why cloud-hosted AI infrastructure providers have become a specific target.

What affected users and customers should do

If you're a Fireworks AI customer, or a customer of a company that uses Fireworks (such as Notion's AI Insights feature), here's what the available guidance suggests:

  • Check for direct notification. Both Fireworks and Notion say they are contacting affected accounts directly. Notion specifically says a workspace that hasn't heard from it has not been identified as affected — though that determination could change as the investigation continues.
  • Rotate any credentials you may have exposed. If you pasted API keys, passwords, or other secrets into a Fireworks-connected product (including Notion AI) around the incident window, rotate them now rather than waiting for a definitive confirmation that they were accessed. This is standard practice after any vendor incident involving a credential-adjacent system — the same advice security teams give after an MFA fatigue attack or other credential-focused intrusion.
  • Watch billing and account activity. Because billing addresses, invoice history, and account-balance data were exposed for Fireworks customers broadly, treat follow-up emails or calls referencing your Fireworks account skeptically — exposed billing details can be reused in targeted phishing.
  • Ask your vendor directly about model weights and datasets. If your organization trained or stored custom models or datasets on Fireworks, Fireworks' bulletin doesn't make clear which specific accounts fall into the "limited set" affected — contacting Fireworks directly through the incident notification or [email protected] is the only way to get an account-specific answer right now.
  • Review third-party AI integrations generally. This incident is a reminder that enabling an AI feature inside a familiar product (like Notion AI Insights) can mean sending your data to a separate infrastructure vendor with its own risk profile — worth reviewing for any tool your organization uses.

As with other recent disclosures where companies gradually widened the scope of what they knew, the practical lesson is that initial breach bulletins can understate eventual scope. It's worth checking back on Fireworks' bulletin and Notion's advisory directly for updates rather than relying on a single snapshot.

Frequently asked questions

What happened in the Fireworks AI data breach?

Fireworks AI disclosed a breach on October 9, 2026 involving stolen credentials. A November 9 update said a second intrusion, via a compromised third-party cloud provider hosting Fireworks' R&D machines, led to wider exposure: billing and account data, resource metadata, and — for a limited set of accounts — secrets, incomplete model weights, and datasets.

Was my credit card number exposed?

Fireworks says no. The company states it does not store payment card numbers, so they were not impacted. Billing and shipping addresses, invoice and payment history, account balance, and credit usage were affected, however.

What does Fireworks mean by 'incomplete model weights'?

Fireworks hasn't explained this in detail. Its bulletin confirms a subset of incomplete model weights and datasets was accessed for a limited set of accounts, but doesn't define what 'incomplete' means or which datasets were involved.

Is the Notion AI Insights exposure the same incident as the Fireworks breach?

It stems from the Fireworks breach. Notion disclosed that AI Insights chat transcripts sent to Fireworks for processing were exposed, affecting fewer than 1% of Notion workspaces. Notion says this was not a breach of Notion's own systems, and most exposed transcripts weren't labeled with an author or workspace.

What should I do if I use Fireworks AI or Notion AI Insights?

Check whether you've received a direct notification from Fireworks or Notion, rotate any credentials you may have pasted into a Fireworks-connected product, and watch for phishing that references exposed billing details. If you stored custom models or datasets with Fireworks, contact the company directly for account-specific confirmation.

Who is behind the attack?

Fireworks has not named the attacker but links the activity to a financially motivated group that resells stolen LLM access, a pattern researchers call 'LLMJacking.' Fireworks says it has worked with Palo Alto Networks' Unit 42 and notified law enforcement.

Sources

More on Fireworks AI →Fireworks AINotiondata breachAI securityLLMJackingcloud security
Sana Qureshi
Written bySana Qureshi

Sana Qureshi runs the security and privacy desk. She reports on actively exploited vulnerabilities, vendor patches and data breaches, and covers the password managers, VPNs and authentication tools readers use to protect themselves. Her alerts cite vendor advisories, CISA and the CVE record directly.

More from Security & Privacy

See all