What happened in the second Fireworks AI breach?
Fireworks AI confirmed on Nov. 9 that a second security incident exposed customer data, a month after it disclosed an earlier credential compromise. The company published the news as a revision to its October security bulletin rather than as a separate advisory.
The timeline in that update is unusually specific. On Oct. 23, a third-party cloud provider that hosted Fireworks research and development machines suffered its own security incident. On Oct. 24, an intruder used that access to reach Fireworks development machines and run an unauthorized workload.
Fireworks says it detected the activity the same day and closed known critical vulnerabilities. By Oct. 25, it had found no further evidence of successful unauthorized access on the platform customers use. The second intrusion came 18 days after the first incident was identified.
The first incident began on Oct. 6. Fireworks found unauthorized use of internal credentials, which exposed environment variables used in evaluation jobs inside a limited number of customer accounts. Its original bulletin said the company had not detected access to customer models, datasets, training data or inference traffic.
That initial assessment did not hold. The Nov. 9 update says later forensics found impact to account and billing information. For a limited set of accounts, Fireworks also found impact to secrets, datasets and incomplete models.
The sequence in the company's own bulletin reads like this:
- •Oct. 6: Fireworks identifies unauthorized use of internal credentials and begins containment.
- •Oct. 7: The company starts notifying affected customers.
- •Oct. 9: Fireworks rotates employee API keys and publishes its first bulletin that evening.
- •Oct. 10: It revokes additional user API keys and rotates service-account tokens after a credential review.
- •Oct. 23: A third-party cloud provider hosting Fireworks R&D machines reports a security incident.
- •Oct. 24: The intruder reaches Fireworks development machines and runs an unauthorized workload; Fireworks detects it the same day.
- •Nov. 9: Fireworks expands the disclosure to billing, account, dataset and partial model-weight impact.
What customer data was exposed?
The billing data includes billing and shipping addresses, invoice and payment history, account balance and credit usage, according to the bulletin. Fireworks says it does not store payment card numbers, so those were not affected.
The company also lists resource metadata such as Fireworks identifiers, display names, descriptions and creation times. That data is less sensitive than a live key, but it can help an attacker map accounts, workloads and internal projects.
The more serious category is customer material. Fireworks says "a subset of incomplete model weights and datasets were accessed" for certain customers. It is notifying those customers separately and giving them recommended response steps.
The bulletin does not explain how incomplete those weights were. Partial checkpoints may not run as a finished model, but they can still contain proprietary training progress. Unlike a password, they cannot be rotated and replaced.
Fireworks says its customer-facing inference service, Training SDK and Nexus platform were not affected. All three are zero data retention by default, meaning Fireworks does not store prompts or training data at rest in those services.
That claim covers the production services, not the development environment the attacker reached. Fireworks has disabled self-serve dataset uploads and downloads for Batch Inference and dataset-based training while it investigates.
The company describes its findings as preliminary and says the investigation may reveal more. Customers should read the bulletin as a status report, not a final scope statement.
The jump from "no evidence" to confirmed dataset and weight access is not necessarily a contradiction. The first review covered a credential leak in evaluation jobs. The second covered a different path through a compromised cloud host and into development machines. Each expanded the forensic scope.
Why is Notion named in the Fireworks AI incident?
Notion published its own notice after Fireworks confirmed that a new dataset included Notion data. The workspace company says fewer than 1% of Notion workspaces may be affected. The exposed material was unattributable chat sessions with Notion AI.
Notion defines unattributable narrowly. The chats were not labeled with the author or workspace, and the transcripts did not include tool call results. Beyond the words in the transcript, Notion says there is no information tying a chat to a workspace, user or company.
The sequence matters. Notion disconnected its systems from Fireworks on Nov. 3, deleted data held in the Fireworks system and halted training runs. On Nov. 6, Fireworks warned that actors may have accessed customer data. On Nov. 8, Fireworks confirmed the dataset.
Notion used Fireworks for classification jobs behind its Insights feature, which helps customers understand their own AI usage. It also used Fireworks as an inference provider for open-weight models inside the product. It has disabled Insights and stopped serving open-weight models through Fireworks.
Affected workspace owners can review the data inside Notion through a security incident review screen. Enterprise plans can download a CSV, while other plans can share a link that lets members see their own affected data. Notion recommends rotating any passwords or API keys that team members pasted into Notion AI during the exposure window.
Notion also confirmed that Learning and Early Access Program customers volunteered data. Notion used that data to experiment with training task-specific models on Fireworks. This is a normal way to test new AI features. It is also how AI vendors end up holding more data than a customer expects.
Unattributable does not mean harmless. A transcript can still contain a customer name, a pricing discussion or a credential someone pasted into a prompt. Notion's own advice is to "treat this unattributable data as if it were exposed."
What is LLMJacking, and why target AI infrastructure?
Fireworks says evidence links the intrusion to a financially motivated cybercriminal group. The group aimed to sell tokens from unauthorized model access through illicit routing gateways. That practice is known as LLMJacking.
Google's Threat Intelligence Group documented that market on Nov. 3. Its researchers found fraudulent charges above $100,000 in a single day. At some organizations, cumulative charges reached $1 million before access was revoked.
The group described "LLM Mixers" that pool stolen credentials, meter usage and resell access. One service, Jet Night, handled more than 500,000 API calls in under eight hours. It passed nearly $1 million in token costs to a victim.
The economics are unusual. Google says Jet Night advertised an unlimited plan for about $60 a month, while comparable consumer access typically costs $100 to $250. Stolen keys make that discount possible.
Google's report also explains how the keys are found. Attackers target exposed or stolen cloud identity tokens, including service-account keys, GitHub personal access tokens and active session tokens. They also harvest API keys from public code repositories.
Validation is designed to stay quiet. A single-token request can confirm that a key still works without triggering volume alerts. Attackers then create and delete short-lived service accounts to stay under rate limits, a pattern Google calls burn and rotate.
Wiz documented the wider pattern in an Oct. 30 report. At least nine AI developer platforms disclosed incidents in October, including Fireworks, Railway, Warp, Lightning AI, Northflank and OpenCode.
Several of those incidents involved credential access rather than one shared software flaw. The common thread was stored environment variables that granted AI access. "Any other month, and this would be the top security story," the researchers wrote.
Wiz also traced downstream abuse of stolen enterprise keys. In two cases, attackers made roughly 800,000 model calls over 30 hours and 1.7 million calls over 13 hours. At list price, Wiz estimated those two cases alone ran above $100,000 in usage.
The report does not tie those resale IP addresses directly to the Black October breaches. It shows that a market exists and that stolen keys move through it, not that every incident had the same operator.
Are Fireworks customers leaving after the breach?
Fireworks is not a small target. The company raised $1.5 billion at a $17.5 billion valuation in July, led by Atreides Management, Index Ventures and TCV, with Nvidia participating. It said annualized revenue had passed $1 billion and that it handled 40 trillion tokens a day.
Fireworks was founded in 2022 by former Meta and Google engineers, and CEO Lin Qiao previously led Meta's PyTorch team. CNBC reported that the company had about 200 employees in July and relied on more than 20 compute suppliers.
It serves more than 10,000 customers, according to its investors and later reporting. Notion and Cursor are among them. CNBC reported that Cursor once supplied about half of Fireworks' revenue before the company diversified. Elastic, GitLab and MongoDB are also customers.
Dealroom reported in October that Business Insider had described early funding talks at Fireworks. Those talks were reported at a valuation of $30 billion to $40 billion. Dealroom noted that the talks were preliminary and that no round size had been disclosed.
Customer reaction has been muted so far. Startup Fortune reported that only a couple of clients had left, largely because replacement GPU capacity is scarce. Moving a tuned model means finding equivalent hardware, re-benchmarking latency and shifting weights without an outage.
That is a market fact, not a security endorsement. Switching costs give an incumbent room to repair trust, even when the incident is serious. The Information framed the risk bluntly in its headline, calling the hack a threat to a $1 billion business.
Fireworks holds SOC 2 Type 2 and ISO 27001, 27701 and 42001 certifications. The certifications cover controls and management systems, but they cannot stop a determined attacker from stealing credentials. The breach is a reminder that certification is not immunity.
What should teams using hosted inference do now?
Fireworks has published concrete steps for customers that received a notice. Replace revoked Fireworks API keys and update every application that used them. Rotate third-party credentials with the providers that issued them, not just inside Fireworks.
The company also asks customers to review logs and data tied to exposed third-party credentials. Check for unrecognized users, evaluators, training jobs and datasets. Fireworks says "Removing a stored secret alone does not invalidate the underlying credential."
Teams that did not receive a notice still have work to do. Keep copies of model weights and datasets under your own keys. Ask vendors which development systems can reach customer data, not only which production systems hold it.
The harder question is architectural. Hosted inference turns a vendor into part of your security boundary. Its evaluation jobs, support tools and R&D machines may hold copies of data that your production contract says is deleted. Ask for that map in writing.
Zero data retention is a useful control, but it is scoped to specific services. Fireworks says inference and training keep no customer data at rest by default. Its development environment held datasets and partial weights, which is where the exposure occurred.
A short checklist covers most of the exposure:
- •Inventory every API key and secret stored in evaluation jobs, notebooks and CI systems, then rotate anything that touched an affected vendor.
- •Scope keys to a single task and a single model, with spending limits and short expiry dates.
- •Watch token spend for sudden spikes, especially overnight and from unexpected regions.
- •Keep an independent copy of weights and datasets, and test restoring them before an incident.
- •Ask vendors which R&D and analytics systems can access customer data, and require notice when that access changes.
- •Price and test a second inference provider, even if you do not plan to switch.
What is still unknown, and what happens next?
Fireworks says Unit 42 is still working on the investigation and that it will provide further updates as appropriate. Law enforcement has been notified. The company has not set a timetable for finishing the review.
Several facts remain undisclosed. Fireworks has not named the third-party cloud provider. It has not said how many customer accounts received notices. It has not described how complete the accessed model weights were, or whether the data has appeared in any resale channel.
Three questions will shape the fallout. Did any stolen weights or datasets surface publicly? Will large customers demand audits or move workloads? And will the October funding talks proceed on the same terms?
The broader trend is unlikely to reverse. Google and Wiz both describe an active market for stolen model access, with brokers, resale sites and automated key checks. Every platform that stores keys for many customers is a candidate.
Fireworks has taken the expected first steps: disclosure, forensics, credential rotation and a public timeline. The harder test is whether it can convince customers that its development environment deserves the same scrutiny as its production stack.
