Disclosure: this article summarises Fireworks AI’s own security bulletins and reporting by BleepingComputer, Bloomberg and other outlets. Details may change as investigations continue. We have no inside knowledge of the incident.
The Fireworks AI breach turned out to be two separate intrusions, three weeks apart. Fireworks, an AI inference and fine-tuning platform backed by NVIDIA, first disclosed on October 9, 2026, that attackers had reached secrets used in evaluation jobs. On November 9 it published a much larger update: a second intrusion, through a third-party cloud provider, had reached billing and account information and, for a limited number of customers, secrets, datasets and unfinished model weights.
The honest lesson is uncomfortable for anyone building on AI platforms. The weak point was not the model and not the customer, but a provider two steps removed. Notion, one of Fireworks’ customers, later confirmed that AI chat transcripts from fewer than 1% of its workspaces were involved. If your product sends data to an AI vendor, the Fireworks AI breach is a reminder that your security now depends on theirs, and on whoever they depend on.
Table of Contents
- The Fireworks AI Breach Timeline
- What the Fireworks AI Breach Exposed
- What the Fireworks AI Breach Did Not Touch
- How the Fireworks AI Breach Reached Notion
- Why AI Inference Vendors Are Targets
- Fireworks AI Breach: Why Stolen Weights Matter
- Seven Lessons From the Fireworks AI Breach
- After the Fireworks AI Breach: What to Do This Week
- Questions to Ask Any AI Vendor
- How the Fireworks AI Breach Was Disclosed
- The Fireworks AI Breach and AI Supply Chains
- FAQ
- Verdict
- Sources and Further Reading

The Fireworks AI Breach Timeline
| Date (2026) | Event |
|---|---|
| October 6 | Fireworks identifies the first intrusion, involving stolen internal credentials |
| October 9 | First security bulletin: secrets used in evaluation jobs were accessed |
| October 23 | A third-party cloud provider hosting Fireworks research and development machines suffers its own security incident |
| October 24 | An unauthorised party reaches Fireworks development machines and runs a workload there |
| October 25 | Fireworks reports no further evidence of unauthorised access |
| November 6 | Fireworks notifies Notion of an incident affecting fewer than 1% of Notion workspaces |
| November 8 | Notion confirms AI chat transcripts for its Insights feature were involved |
| November 9 | Fireworks publishes its expanded bulletin; Bloomberg reports the second incident |
Let’s Data Science made a sharp observation about the framing. Fireworks describes everything as a single October 2026 security incident, but its own bulletins show two separate intrusions, three weeks apart. That distinction matters, because it means the first response did not prevent the second.
FourWeekMBA’s headline captured the market reaction: a company valued at $17.5 billion hacked again 18 days after its first breach. Investigators from Palo Alto Networks’ Unit 42 are involved, and law enforcement has been notified, according to AI TLDR’s summary of the bulletin.
What the Fireworks AI Breach Exposed
The November 9 update widened the picture considerably. According to Unite.AI and Fireworks’ bulletin, the attackers accessed:
- Account and billing information for customers.
- Secrets, such as keys and credentials, for a limited group of customers.
- Datasets that some customers had uploaded, for example for fine-tuning.
- Incomplete model weights, meaning unfinished fine-tuned models, for a small subset of customers.
Fireworks said attackers could not access payment card numbers because it does not store them. That is a standard reassurance and a real one; card data is the most heavily regulated category, and not holding it limits one class of harm.
The secrets are the most urgent problem for affected customers. A stolen API key or cloud credential can be used to run up bills, reach other systems or pull more data. Anyone notified that secrets were involved should rotate them immediately, not after reading the rest of the bulletin.
What the Fireworks AI Breach Did Not Touch
Fireworks says it has not detected any impact to its Inference service, its Training SDK or Nexus, which run with zero data retention by default. In other words, prompts and responses passing through its main inference product were not stored, so there was nothing there to steal.
That design choice is the best defensive lesson in the whole episode. Data that is never retained cannot leak. The exposed categories were the ones that, by their nature, have to be stored: billing records, uploaded datasets, fine-tuning outputs and secrets.
It also explains why the Notion data was different. Notion’s AI chat transcripts were reportedly used for a specific feature, Insights, which meant they were stored for processing rather than passed through and discarded.
How the Fireworks AI Breach Reached Notion
The downstream effects are what turned a vendor incident into a broader story. On November 6, Fireworks told Notion that attackers might have accessed real customer data. On November 8, Notion confirmed that AI chat transcripts used for its Insights feature were involved, touching fewer than 1% of Notion workspaces.
The chain then extended further. Breached.company reported that Notion notified at least one of its own customers, Render, that its workspace was among those affected. So a breach at a cloud provider reached an AI vendor, then a productivity company, then that company’s customers. Four layers, each depending on the one beneath it.
A Fireworks spokesperson told Bloomberg the company was actively working with customers including Notion and Cursor to manage the incident. Cursor had separately published its own incident report on November 13, which our coverage of the Cursor security incident discusses.
Why AI Inference Vendors Are Targets
AI platforms hold an unusually attractive mix of assets. They store customers’ API keys and cloud credentials. They hold proprietary datasets that companies uploaded for fine-tuning. They hold model weights, which can represent months of work. And they control large amounts of expensive GPU compute.
The compute alone is a motive. Explainx linked the Fireworks AI breach to the wider pattern of LLMjacking, in which attackers hijack access to AI services and either resell it or use it themselves. Google’s threat intelligence group reported in November that such attacks have left some organisations with bills of more than $100,000 a day. An inference platform with access to many customers’ accounts is a high-value target for that kind of operation.
TechJack Solutions noted that Fireworks attributed the first incident to a financially motivated actor. That fits the pattern: most attacks on AI infrastructure so far look like theft for profit rather than espionage.
Fireworks AI Breach: Why Stolen Weights Matter
Unfinished model weights may sound less alarming than customer data, but they deserve attention. A fine-tuned model encodes what it was trained on. Depending on the data and the method, it can sometimes be coaxed into revealing fragments of its training examples.
For a company that fine-tuned a model on internal documents, support transcripts or code, stolen weights are a potential leak of that material, not just a loss of intellectual property. The incomplete state of the weights may limit their usefulness, but it does not remove the risk entirely.
There is also the competitive angle. A custom model can represent a real head start. Even partially trained weights could save a competitor time. Affected customers should assume their tuned models are no longer exclusive and plan accordingly.
Seven Lessons From the Fireworks AI Breach
- Your vendor’s vendors are your risk. The second intrusion came through a cloud provider. Ask who your AI vendor depends on.
- Zero retention works. The services that kept nothing lost nothing. Prefer zero-retention settings wherever your use case allows.
- Minimise uploaded data. Datasets uploaded for fine-tuning were exposed. Upload only what training needs, strip personal data first and delete datasets after use.
- Treat secrets as short-lived. Keys stored with a vendor should be scoped narrowly and rotated regularly, so a theft has a short shelf life.
- Read the second bulletin. The first disclosure understated the impact. Follow incidents until they are closed.
- Contract for downstream notification. Notion learned of the impact weeks after the intrusion. Breached.company recommended reviewing contracts for explicit downstream-notification obligations, which is sound advice.
- Know what you sent where. You cannot assess exposure without a map of which data goes to which AI service.
After the Fireworks AI Breach: What to Do This Week
Whether or not you used Fireworks, the Fireworks AI breach is a good prompt for a quick review.
- List every AI vendor your systems call, including those embedded in other tools.
- For each, note what data you send, whether it is retained and for how long.
- Rotate API keys and credentials stored with any vendor that has had an incident, and set calendar reminders to rotate the rest.
- Turn on zero-retention or no-training settings wherever available.
- Delete fine-tuning datasets and old model checkpoints you no longer need.
- Set spending alerts and hard caps on AI accounts, which limit the damage from stolen keys.
- If you were notified as an affected customer, follow the specific instructions in your notice first; Fireworks asks customers to reply to the notification email or contact its security team.
Questions to Ask Any AI Vendor
Procurement reviews for AI services often focus on model quality and price. After the Fireworks AI breach, a few security questions belong on the list.
- Which data do you retain, for how long, and can retention be switched off?
- Which third-party providers host your infrastructure, and how are they assessed?
- How are customer secrets stored, and can we use short-lived or scoped credentials instead?
- How quickly will you notify us of an incident, and will you tell us if our own customers’ data is involved?
- Can we delete uploaded datasets and model weights on demand, and get confirmation?
- What independent security certifications or audits cover your platform?
Good vendors will have clear answers. Vague answers are information too.
How the Fireworks AI Breach Was Disclosed
The way an incident is communicated shapes how much damage it does, and the Fireworks AI breach offers lessons on both sides.
On the positive side, Fireworks published security bulletins on its own site rather than relying only on private emails, named the categories of data involved, said what was not affected and explained the architectural reason, zero data retention, for the services that escaped. It also brought in an outside incident response team and notified law enforcement. Those are the basics of a credible response.
On the other side, the first bulletin on October 9 described a narrower incident than the one customers eventually learned about. The second intrusion happened on October 24, but the expanded disclosure came on November 9, more than two weeks later. Notion was told on November 6. For customers trying to decide whether to rotate credentials or warn their own users, those gaps matter.
Describing two intrusions as a single incident is also a choice worth questioning. It is understandable from a communications point of view, since both may trace back to the same broad campaign. But customers assess risk differently when they know a second attack succeeded after the first was discovered. Clear separation of the two would have helped.
The broader point applies to every vendor. Speed and completeness of disclosure are part of security. A vendor that tells you early, even with incomplete information, gives you time to act. One that waits for a full picture may be more accurate and less useful.
If you are writing your own incident plan, the Fireworks AI breach suggests a simple rule: notify early with what you know, update on a fixed schedule, and say explicitly when the scope has grown. Customers forgive bad news far more readily than they forgive surprises.
It is also a reminder to keep contact details for your vendors’ security teams somewhere you can find them quickly. When an incident notice arrives, the first useful question is usually whether a specific key, dataset or model of yours was in scope, and getting that answer fast depends on knowing whom to ask.
The Fireworks AI Breach and AI Supply Chains
The Fireworks AI breach is part of a wider autumn of incidents affecting AI developer tools. Breached.company’s review recommended checking contracts with platforms including Northflank, Trigger.dev, Chainstack and WorkOS, all named in recent incident reporting, and BleepingComputer separately reported that WorkOS said an attacker used the ARTEX framework to breach an internal sales app.
The common thread is that AI products are now assembled from many services: model hosts, vector databases, agent frameworks, observability tools and cloud providers. Each one is a dependency, and each dependency is a potential entry point. Traditional software has lived with this for years. AI stacks are newer, move faster and often skip the vendor reviews that older software went through.
That is likely to change. Expect enterprise buyers to demand security questionnaires, audit reports and notification clauses from AI vendors as standard, and expect insurers to start asking about AI supply chains too. Vendors that can show zero-retention designs, strong secret handling and fast, honest disclosure will win business from those that cannot.
For smaller teams without a procurement department, the practical version is simple: fewer vendors, less data sent, shorter-lived credentials and a written list of what goes where.
Frequently Asked Questions About the Fireworks AI Breach
What happened in the Fireworks AI breach?
Fireworks disclosed two intrusions in October 2026. The first, detected on October 6, involved stolen internal credentials and evaluation-job secrets. The second, on October 24, came through a compromised third-party cloud provider and reached customer data.
What data was exposed?
Account and billing information, and for a limited number of customers, secrets, datasets and incomplete model weights. Fireworks says it does not store payment card numbers.
Was inference traffic affected?
Fireworks says it detected no impact to its Inference service, Training SDK or Nexus, which run with zero data retention by default.
Which customers were affected?
Fireworks named Notion and Cursor among customers it was supporting. Notion said fewer than 1% of its workspaces were affected, involving AI chat transcripts for its Insights feature.
Who is investigating?
Palo Alto Networks’ Unit 42 is investigating, and law enforcement has been notified, according to the bulletin summary.
What should I do if I use Fireworks?
Follow any notification you received, rotate credentials stored with the platform, delete datasets and checkpoints you no longer need, and contact Fireworks’ security team with questions.
Verdict
Use the Fireworks AI breach as a prompt to map and shrink your AI data footprint. The services that retained nothing lost nothing, while stored secrets, datasets and weights were exposed. That is the clearest guide to what to change.
The honest caveat is that investigations are ongoing and the picture has already grown once. Initial disclosures understated the impact, and downstream effects reached customers’ customers weeks later. Details here may be revised.
Next step: this week, list every AI vendor you use, rotate the credentials you store with them, switch on zero-retention settings and add downstream-notification terms to your next contract renewal.
Sources and Further Reading
- Fireworks AI: incident update, the company’s own bulletin.
- BleepingComputer: new breach after cloud provider hack, for the second intrusion.
- Let’s Data Science: the two intrusions, for the detailed timeline and Notion impact.
- AI TLDR: breach widens, for what was and was not affected.
- Breached.company: downstream spread, for contract advice.
On this site: read about the Cursor security incident, understand the ARTEX AI warning, protect agents with prompt injection and agent security, and compare self-hosting options in LLM inference servers compared.

