Subscribe
0%
Illustration for: Fireworks AI's breach widens: billing data, some datasets and unfinished model weights taken. What customers should do

Fireworks AI's breach widens: billing data, some datasets and unfinished model weights taken. What customers should do

Fireworks AI now says a second intrusion on October 24, through a cloud provider hosting its R&D machines, reached billing records for customers and, for some accounts, secrets, datasets and incomplete model weights. Its inference service shows no impact, but anyone who trained or evaluated models there should check their notice and rotate keys.

Listen to this article0:00 / 7:53

Fireworks AI, the model-hosting and fine-tuning platform, says its October security incident reached far more customer data than it first reported. In an updated security bulletin dated November 9, the company says an attacker got onto its development machines on October 24 and ran an unauthorized workload. That exposed billing and account information and, for a limited set of accounts, stored secrets, parts of datasets and incomplete model weights.

This matters because Fireworks is where many companies bring proprietary data to fine-tune and evaluate models. Its first disclosure, in early October, said only evaluation-job secrets were affected and that there was no sign of access to models or datasets. The update reverses that for some customers. It's also the second intrusion in under three weeks. Bloomberg's headline, as cited by RuntimeWire, calls it the company's second hack in weeks.

Two incidents, one bulletin

Date (2026) What Fireworks says happened
Oct 6 Activity identified and containment begins: internal credentials misused to read environment variables in evaluation jobs
Oct 7–9 Affected customers notified, employee API keys rotated, first bulletin published
Oct 10 More user API keys revoked, service-account tokens rotated
Oct 23 A third-party cloud provider hosting Fireworks R&D machines has its own security incident
Oct 24 Attacker uses that access to reach Fireworks development machines; detected the same day
Oct 25 No further evidence of successful access to the customer platform, Fireworks says
Nov 9 Updated bulletin widens the list of affected data

Fireworks doesn't name the cloud provider and doesn't say how many customers are affected. It says affected customers are being notified directly.

What was exposed

According to the update:

  • Resource metadata — resource IDs, display names, descriptions and creation times.
  • Billing information — billing and shipping addresses, invoice and payment history, account balance and credit usage. Fireworks says it doesn't store card numbers, so they weren't exposed.
  • Secrets for a limited set of accounts.
  • A subset of datasets and incomplete model weights for certain customers.

Fireworks says it found no impact on its Inference, Training SDK or Nexus services, which it says are zero data retention by default. It warns that the investigation is ongoing and more may come to light.

Who did it, and why

Fireworks links the activity to a financially motivated group that sells access to stolen LLM capacity through illicit routing gateways, a practice it calls "LLMJacking." That's the company's attribution; it doesn't name a group. The motive suggests the attackers were after credentials and compute as much as data. That makes stored API keys for other AI providers the most urgent thing to deal with.

What Fireworks has changed

The company says it told law enforcement and brought in Palo Alto Networks' Unit 42. It also says it replaced credentials and service accounts, removed unneeded admin privileges, isolated affected systems and tightened storage access and monitoring. It has turned off user-initiated dataset uploads and downloads for Batch Inference and dataset-based training while the investigation continues; Training SDK stays available, according to the bulletin and Superpower Daily's summary. If a pipeline moves datasets in or out of Fireworks on its own, expect it to fail until the transfers come back.

What to do now

The November update tells affected customers to follow the steps in their notification email. Fireworks' original October bulletin gives more concrete guidance, and the same steps make sense here:

  1. Find your notice. Search your inbox and your billing contact's inbox for Fireworks incident emails from October and November. The October bulletin says customers who got no notice don't need to take incident-specific action. Given the expanded scope, it's still worth checking again.
  2. Replace any revoked Fireworks API keys listed in your notice, and update your apps. Old keys will fail authentication.
  3. Rotate third-party credentials you stored as Fireworks secrets: OpenAI, Anthropic, cloud, Hugging Face and database keys. Do this with the issuing provider. Fireworks points out that deleting the stored secret doesn't revoke the credential.
  4. Check those providers' logs for unexpected usage. Given the LLMJacking motive, look for spikes in token spend or calls from unfamiliar IPs. Fireworks says to rotate even if nothing looks wrong.
  5. Audit your Fireworks account for unfamiliar API keys, users, evaluators, training jobs or datasets, and report any to Fireworks.
  6. Assume uploaded datasets may be exposed if your notice mentions datasets or models. Work out what was in them (customer records, proprietary text, personal data) and whether that triggers your own notification duties.
  7. Watch for invoice-themed phishing. Billing addresses, invoice history and balances were taken, which makes convincing fake "Fireworks billing" emails easy to write.

Questions go to security-notice@fireworks.ai or a reply to your notification email, the company says.

Update, November 11: Notion AI chats were in the stolen data

Notion says some Notion AI chat transcripts were in a dataset taken in the Fireworks incident, affecting fewer than 1% of its workspaces, according to its help-center notice as reported by Tech Bytes and Cryptika. Notion used Fireworks to run classification jobs for Insights, its AI-usage analytics feature, and to serve open-weight models.

Date (2026) What Notion says happened
Nov 2 Fireworks alerts Notion; an initial review finds only synthetic test data
Nov 3 Notion disconnects from Fireworks, deletes its data there and halts training runs
Nov 6 Fireworks warns that customer data may have been accessed
Nov 8 Fireworks confirms the dataset; Notion finds Insights chat transcripts in it

Notion describes the transcripts as "unattributable": not labelled with a user or workspace, and without tool-call results such as page content. It says no other workspace data was affected. But whatever people typed into a chat is still exposed, so Notion recommends treating the transcripts as exposed and rotating any passwords or API keys pasted into Notion AI. Insights is disabled, Notion no longer serves models through Fireworks, and it has halted training experiments that used volunteered data from its early-access program.

Workspace owners can review affected data from ? → Security incident review in the sidebar, on desktop or web (not mobile). On Enterprise plans with managed members, only the owner sees the full CSV. On other plans, members download only their own data. Notion says workspaces that haven't been contacted aren't affected.

Update, November 14: Uber drivers' licence data was taken too

Uber is notifying drivers that their personal information, including driver's licence details, was stolen through Fireworks, TechCrunch reported on November 13. An Uber spokesperson tied the notices to a security incident at "one of our third-party vendors, Fireworks.ai." This week Uber filed notices with several US state attorneys general, California's among them.

State People notified (per filings)
South Carolina 3,430
Massachusetts 2,567
Rhode Island 1,873
Illinois Some residents; count pending
California Notified; no count given

That's at least 7,870 people across three states, by Gizmodo's count. Uber hasn't given a national total, and TechCrunch expects the number to rise as more state filings appear. Gizmodo's timeline from the notices: Fireworks told Uber on October 9 that stolen internal credentials had been used to get into Uber's account with it, and Uber confirmed on October 13 that its data had been accessed. That fits the first, October intrusion described above. Uber says it cut its data connections to Fireworks and hasn't seen any misuse of the information so far.

If you drive for Uber: read the notice letter. Uber is offering free credit monitoring and fraud help through Cyberscout, and Gizmodo reports drivers have 90 days from the letter date to enroll. Uber also recommends checking account statements and credit reports. A stolen licence number is useful for identity fraud, so consider a credit freeze with the three bureaus. Treat calls or emails that mention your Uber account or licence as possible phishing.

What to watch

The key open questions are how many accounts lost datasets or weights, which cloud provider was breached, and whether other AI platforms that use the same provider were hit too, and how many more Fireworks customers, like Notion and Uber, turn out to have passed personal data through it. Fireworks says it will update the bulletin as the investigation goes on.

Sources

  1. Fireworks AI — Update on October 2026 security incident fireworks.ai
  2. Fireworks AI — Fireworks October 2026 security incident fireworks.ai
  3. RuntimeWire — Fireworks AI says cloud-provider breach exposed customer billing data and model weights runtimewire.com
  4. Superpower Daily — Fireworks discloses wider breach impact, including customer datasets and unfinished models superpowerdaily.com
  5. Tech Bytes — Fireworks AI security breach impacts Notion and their customers techbytes.app
  6. Cryptika — Notion Confirms AI Chat Transcripts Exposed in Fireworks AI Breach cryptika.com
  7. TechCrunch — Uber confirms hackers stole thousands of drivers' license information in data breach techcrunch.com
  8. Gizmodo — Hackers stole thousands of Uber drivers' data through an AI startup gizmodo.com

#Fireworks AI#data breach#LLMjacking#API keys#model weights#AI infrastructure