Skip to main content
All incident reports
Incident report

Fireworks AI breach exposes customer billing data and Notion AI chat transcripts

Fireworks AI says two intrusions in October exposed customer billing data, secrets and incomplete model weights, and has linked the activity to financially motivated criminals reselling access to large language models. Notion says transcripts from its AI Insights feature were among the data exposed, affecting fewer than 1% of its users.

· First published

The short version

  • The breaches. Fireworks AI says it found unauthorised use of internal credentials on 6 October, and a second intrusion through a third-party cloud provider on 23-24 October that exposed more customer data.
  • What was taken. The data includes resource metadata, billing and shipping addresses, invoice and payment history, account balance and credit usage, plus incomplete model weights and datasets for some customers. Fireworks says payment card numbers were not affected.
  • Notion's exposure. Notion says the exposed dataset held Notion AI chat transcripts from its Insights feature. It says fewer than 1% of its users are affected and has disabled the feature and stopped using Fireworks to serve open-weight models.
  • Attribution. Fireworks says the evidence links the activity to a financially motivated criminal group that sells access to stolen LLM capacity. Mandiant calls this LLMJacking and says it has handled several such incidents in 2026.

Timeline

DateEvent
6 OctoberFireworks detects unauthorised use of internal credentials at 9 a.m. PDT.
10 OctoberFireworks publishes a notice saying environment variables, API keys and tokens were accessed in a limited number of accounts.
23 OctoberAttackers reach Fireworks' development environment through a third-party cloud provider hosting its R&D machines.
24 OctoberAn unauthorised workload run on Fireworks development machines leads to further customer data exposure.
25 OctoberFireworks says it has found no further evidence of successful unauthorised access.
2 NovemberFireworks alerts Notion to an exposed dataset that contained only synthetic test data, according to Notion.
3 NovemberNotion disconnects from Fireworks, deletes its data there and halts training runs.
6 NovemberFireworks tells Notion that malicious actors may have accessed customer data, according to Notion.
8 NovemberFireworks gives Notion the dataset it analyses and, Notion says, confirms that customer data including Notion's was accessed.
10 NovemberFireworks publishes an updated notice describing the wider impact.
11 NovemberNotion discloses that its AI chat transcripts from Insights were exposed.

The two October intrusions

Fireworks AI, which lets developers run and fine-tune open models, detected unauthorised use of internal Fireworks credentials on 6 October at 9 a.m. PDT. In a notice published on 10 October it said the intruder had reached environment variables used in evaluation jobs in a limited number of customer accounts, including API keys and tokens stored as secrets. At that point it said it saw no sign that customer models, datasets, training data, inference traffic or personal information had been accessed.

An updated notice published on 10 November described a second intrusion. On 23 October attackers reached Fireworks' development environment through an unnamed third-party cloud provider that hosted its research and development machines; the following day they used that access to run an unauthorised workload that led to further customer data exposure. Fireworks said it had found no further evidence of successful unauthorised access as of 25 October.

The exposed data includes resource identifiers, display names and descriptions; billing and shipping addresses, invoice and payment history, account balance and credit usage; and a subset of incomplete model weights and datasets belonging to certain customers. Fireworks says it does not hold payment card numbers and that they were not affected. It has hired Palo Alto Networks' Unit 42 to investigate and says it revoked and rotated credentials, isolated systems, tightened access controls and improved logging and monitoring.

Notion and downstream exposure

Notion used Fireworks to serve open-weight models inside its product and to analyse stored conversations for Insights, a feature that shows workspace administrators how their teams use AI. Notion says Fireworks alerted it on 2 November to an exposed dataset containing only synthetic test data; by 3 November Notion had cut its connection to Fireworks, deleted its data from the vendor's systems and halted training runs.

Notion says Fireworks warned it on 6 November that malicious actors may have accessed customer data, and gave it the dataset it analysed on 8 November. That dataset, Notion says, included Notion AI chat transcripts from Insights; most were unattributable — not labelled with an author or workspace — and contained no tool call results. Notion says Fireworks had contractually committed to zero data retention for Notion data.

Notion says fewer than 1% of its users are affected, describes the share elsewhere as fewer than 1% of workspaces, and TechCrunch reports it as around 1% of customers. Workspace owners can review the data through Security incident review in Notion's desktop web interface and share a link that lets members download their own transcripts. Notion tells anyone who pasted passwords, API keys or other credentials into an affected chat to rotate them, and says users who received no notice most likely need do nothing.

LLMjacking and resale

Fireworks says the evidence points to a financially motivated cybercriminal group that sells tokens obtained through unauthorised access to large language models to illicit routing gateways. Google Threat Intelligence Group and Mandiant, publishing research the week before Fireworks' update, call the tactic LLMJacking: hijacking an organisation's model access through stolen credentials so the victim pays for someone else's usage.

Mandiant says it has responded to multiple LLMJacking incidents in 2026 across cloud and LLM providers, affecting organisations in North America, South America and Asia Pacific in financial services, retail, technology and transportation. It says fraudulent charges can exceed $100,000 in a single day and reach $1 million cumulatively at some organisations before access is revoked, far faster than the gradual consumption seen in cryptojacking. GTIG describes several clusters, some taking initial access and handing it to others that resell it.

What is not known yet

  • Which criminal group carried out the intrusions; Fireworks has not named it, and Mandiant describes several separate clusters.
  • How many Fireworks customer accounts were affected, beyond the company's statement that it was a limited number.
  • Whether the same actor was behind both the early October and the 23-24 October intrusions.
  • Which third-party cloud provider hosted the research and development machines used as the entry point.
  • How Notion AI transcripts came to be held by Fireworks, given Notion says the vendor had committed to zero data retention for its data.

What to do

If you use Fireworks AI

  1. Check invoices, billing details and credit usage for consumption you did not incur, and report anomalies to Fireworks.
  2. Rotate any Fireworks API keys and service account tokens you have not already replaced, and revoke keys you no longer need.
  3. Follow the instructions in any notice Fireworks sends, since it says it is contacting affected customers.

If you use Notion AI

  1. Open Security incident review in Notion's desktop web interface to see whether your workspace data is listed, and share the download link with members so they can retrieve their own transcripts.
  2. Rotate any passwords, API keys or other credentials that were pasted into Notion AI chats.
  3. Take no further action if neither you nor your workspace owner has received a notice, Notion says; you are most likely not in the affected subset.

For everyone else

  1. Treat anything sent to a third-party AI assistant as data that provider may hold and could expose, and keep credentials and customer data out of prompts.
  2. Review which external AI providers hold your data and what they retain; do not assume a zero-retention clause prevents exposure.
  3. Watch model and cloud spend for sudden spikes, which Mandiant says can exceed $100,000 in a day under LLMJacking.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.