Fireworks AI October 2026 Breach: Billing Data, Model Weights and LLMjacking Update

Fireworks AI widened the scope of its October 2026 security incident on November 9, 2026, saying billing records joined secrets and model artifacts on the impact list and tying the case to financially motivated LLMjacking. The November 9 bulletin update builds on the October 9 notice and names Palo Alto Networks Unit 42 as the forensics partner.

If you route production traffic through Fireworks, treat this as a supply-chain rotation event, not a finished story. Fireworks says the investigation continues and may reveal more data categories.

Update (November 9, 2026): Forbes customer fallout reporting

Forbes reporter Anna Tong published additional customer reaction on November 9, 2026, alongside the bulletin. We could not open Forbes directly in automated fetches (HTTP 403), so the following summarizes lines attributed to Forbes in Techmeme's link-out and secondary republications that cite Forbes by name.

Forbes describes Fireworks as infrastructure used to build specialized models on proprietary data, naming enterprise customers including Uber, Shopify, Upwork, DoorDash, and startups including Cursor and Cognition in its reporting. Treat that customer list as press attribution, not a confirmed victim roll call — Fireworks has not published an equivalent named list in its bulletins.

On customer movement, Forbes says it spoke with two customers who had substantially moved business off Fireworks, with one characterizing a series of very bad hacks of increasing frequency and criticizing communications around the attacks. Two other customers Forbes contacted said they were not aware of any Fireworks hacks — a reminder that migration stories and unaware teams can coexist in the same supplier breach.

Forbes frames the risk as concentration: bespoke training and dataset uploads turn inference vendors into custodians of proprietary weights and customer data, which makes LLMjacking financially attractive. That narrative aligns with Fireworks' own November 9 description of financially motivated actors reselling unauthorized model access, but Forbes adds procurement pressure — teams re-evaluating whether a fine-tuned model host should hold billing addresses, partial checkpoints, and evaluation secrets on the same plane as API keys.

If your security lead is drafting an exit memo, pair Forbes' migration anecdotes with Fireworks' direct notification status for your account. Anecdotes are not a substitute for whether your tenant received a breach letter or saw anomalous credit usage.

Update (November 10, 2026): Notion, Cursor, and named enterprise customers in supplier fallout

Bloomberg Law reported on November 9, 2026 that a Fireworks spokesperson said the company is actively working with customers to manage the incident, including Notion and Cursor, among others. The story situates the widened November 9 bulletin as Fireworks' second known hack in weeks, after the October credential path and the October 23–24 cloud-provider chain described in the official update.

Fireworks' own November 9 security bulletin does not name Notion, Cursor, Harvey, or other logos. Treat Bloomberg Law's customer list as vendor-spokesperson attribution, not confirmation that a specific workspace product was exfiltrated. Fireworks continues to describe impacts as account-specific with direct notification.

For Notion teams, the practical read is indirect risk, not a confirmed Notion SaaS breach in the sources we checked:

• If your org fine-tuned or evaluated models on Fireworks and stored datasets, incomplete weights, or evaluation secrets there, you belong in the limited set Fireworks flagged for deeper impact.
• Zero-data-retention inference paths are still described as unaffected in Fireworks' bulletin; that does not cover artifacts uploaded for training or batch jobs.
• Productivity data inside Notion pages is a separate trust boundary unless your integration copied content into Fireworks-hosted datasets—verify data flows, not headlines.

Teams shipping coding agents should read Cursor's mention the same way: any Fireworks API keys or evaluation jobs in that supply chain need rotation and scope review per Fireworks' October notice playbook, even if day-to-day Cursor chat never touched Fireworks.

Update (November 10, 2026): Notion customer email describes AI chat transcripts

A Hacker News Tell HN post quotes Notion's customer email saying Notion AI chat transcripts from Insights classification appeared in Fireworks breach data, affecting under 1% of workspaces. Notion reports Insights and open-weight Fireworks routing are disabled pending investigation.

Full explainx.ai coverage: Notion AI chats and the Fireworks breach.

TL;DR

What Fireworks says changed in the November update

The bulletin opens by restating the October timeline, then adds categories that were not in the first public summary.

Billing information now sits on the affected list: billing and shipping addresses, invoice and payment history, account balance, and credit usage. Fireworks repeats that payment card numbers are not stored and were not part of the breach.

Model weights and datasets appear for certain customers as incomplete weights and datasets accessed—not necessarily production-ready checkpoints, but still sensitive if your fine-tunes or proprietary data lived in Fireworks storage.

Resource metadata was already in scope: Fireworks resource identifiers, display names, descriptions, and creation timestamps.

The company is notifying impacted customers individually with recommended response actions. If you did not receive mail, that does not prove zero impact; confirm with your account team if you store custom artifacts on the platform.

LLMjacking: why inference vendors are targets

Beads leaking through a crack in a storage vessel, suggesting data exfiltration

Fireworks says evidence points to a cybercriminal threat actor group that sells tokens from unauthorized access to LLMs through illicit routing gateways—behavior security researchers call LLMjacking.

That framing matches Google Threat Intelligence Group and Mandiant research published November 2, 2026, which describe resellers turning stolen cloud and model API keys into cheap inference for buyers. explainx.ai covered the pattern in Google GTIG on LLMjacking: single-day charges above $100,000 and lifetime totals approaching $1 million appear in their case studies.

For builders, the lesson is structural. Keys on an inference host are money in the wire. Attackers do not need your source code if they can bill your credits or resell access. The same economics drove October's wave against AI developer platforms that Wiz summarized as Black October—Fireworks listed alongside Railway, Zeabur, Warp and Northflank in that thread.

Fireworks is also where Jensen Huang spoke at Fireworks Forge days before this update; our Forge coverage focused on models, but supply-chain trust for inference keys matters just as much for teams routing Kimi K3 and Fable routing studies through the platform.

The October 23 cloud-provider chain

The November bulletin adds a second intrusion path that did not appear in the earliest public summary.

On October 23, a third-party cloud provider hosting Fireworks R&D machines experienced a security incident. On October 24, Fireworks says an unauthorized party used that compromise to reach Fireworks development machines and run an unauthorized workload that led to additional customer data impact.

Fireworks reports it detected activity the same day, followed incident response, took immediate remediation, and closed known critical vulnerabilities. As of October 25, it says it had not detected further successful unauthorized access on the customer-facing Fireworks platform.

Read this as nested supply chain risk. Your vendor's vendor's breach can become your dataset exposure even when your own perimeter looked clean. That pattern also showed up when attackers reportedly moved from Fireworks keys into Cursor in commentary on Alex Stamos's Decoder interview—a chain we treat as reported, not independently verified here, but consistent with how infostealers hunt AI credentials.

What Fireworks says was not impacted

Fireworks draws a bright line around its zero data retention (ZDR) products:

• Inference
• Training SDK
• Nexus

It says these paths have not detected impact, with the caveat that the investigation is ongoing.

Operational changes while investigating include disabling self-serve dataset uploads and downloads used by batch inference and dataset-based training. Training SDK remains available with ZDR (no training data stored at rest). Fireworks Lab-based training runs in an isolated environment with extra network and storage protection, per the bulletin.

If your architecture depends on batch jobs that upload datasets through the UI, expect friction until Fireworks re-enables those flows.

Hardening measures Fireworks lists

Working with Unit 42, Fireworks says it implemented:

• Access and credential lockdown — revoked and replaced credentials and service accounts, removed unnecessary elevated privileges, tightened controls on new credential creation.
• Network isolation — restricted unauthorized network paths and isolated affected systems.
• Data protection — reviewed storage and analytics access; enforced controls against public or unauthorized data exposure.
• Forensic and audit integrity — secured and enhanced logging and monitoring.

The bulletin promises 24/7 monitoring and further updates as appropriate.

What developers should do this week

Vault door with a key beside it, illustrating API key rotation and secrets storage

Rotate secrets first. Any Fireworks API key that existed before October 25 should be replaced, with old keys revoked—not disabled, revoked. If you mirrored keys into CI, local .env files, or agent configs, rotate all copies the same day.

Scope and cap keys. Set per-key spend limits and restrict models and endpoints to what each environment needs. LLMjacking succeeds when one stolen admin key can proxy unlimited frontier traffic. AgentBeam's LLMjacking guide walks through budget caps and gateway controls that slow resale abuse.

Audit billing and credits. Because billing addresses and usage history were in scope, watch for unexpected credit draws or invoices to addresses you did not authorize. Pair Fireworks logs with your own gateway logs if you front the API.

Treat custom weights as sensitive. If you trained or uploaded checkpoints, assume incomplete copies could have been read if Fireworks notified you. Do not rely on obscurity for unreleased models; rotate signing secrets and rebuild from trusted bases if your threat model includes competitor intelligence.

Segment inference from data lakes. Keep dataset storage off the same keys that power public-facing agents. Patterns from Northflank and GitBook show control-plane secrets becoming inference abuse within hours.

Comparing the October 9 and November 9 bulletins

Builders auditing vendor communications should read the two bulletins side by side. The October notice already described law-enforcement notification and Unit 42 engagement. It emphasized that Fireworks stored secrets used in evaluation jobs were accessed first. The November update is explicit that the forensic picture grew: billing categories and incomplete customer weights joined the list.

Fireworks does not publish a victim count or a list of affected model IDs. That is normal for an active investigation, but it means your risk call depends on direct notification and on whether you used features that store data at rest—custom training uploads, saved datasets, or long-lived evaluation secrets—versus pure ZDR inference.

Batch inference with uploaded datasets is the workflow most likely to intersect the disabled upload and download paths. If you moved batch jobs to another host this week, keep Fireworks keys revoked until you confirm account status. Teams that only called hosted models through stateless HTTP may have lower exposure, but rotation remains cheap insurance.

Credit usage is worth a separate audit. LLMjacking resellers monetize inference, not GitHub repos. A stolen Fireworks key can burn prepaid credits or post-pay invoices before you see an anomaly. Pair vendor mail about billing fields with your own per-key dashboards and alerts. The same discipline applies if you also hold keys on OpenRouter or other aggregators that front multiple providers.

How this fits November's breach picture

Our running November 2026 cyberattacks list already tracks AI provider and agent incidents as they land. Fireworks belongs next to credential-theft chains that start at vendors, not at your app repository.

The UK's NCSC State of the Cyber Nation report this month also warns that AI-enabled crime rises alongside ransomware pressure on critical infrastructure. LLMjacking is the low-friction version: no ransomware note, just someone else spending your inference budget.

For teams comparing hosts after the incident, see how Fireworks Ember-1 pricing and routing studies assume a trusted API boundary—and rebuild that boundary after any supplier bulletin like this one.

Enterprise procurement questions to ask Fireworks

If your security team is writing a vendor questionnaire this week, these items map directly to the November bulletin language:

• Which SKUs are ZDR by default? Confirm Inference, Training SDK, and Nexus paths match your contract SKU list.
• Are dataset uploads still disabled? Batch pipelines may need a temporary host until Fireworks reopens uploads.
• Who received breach notifications? Align legal and security distribution lists so incident mail is not trapped in a single engineer inbox.
• What metadata left the platform? Resource identifiers can aid reconnaissance even without weights.

Document answers in your vendor risk register next to Northflank and GitBook entries so auditors see a pattern, not a one-off.

Related reading

• Google GTIG: LLMjacking and stolen cloud AI access
• November 2026 cyberattacks tracker
• Railway and OpenRouter key theft checklist
• Jensen Huang at Fireworks Forge
• Fireworks Kimi K3 and Fable routing study
• Hugging Face on open-hub security tradeoffs
• Alex Stamos on AI supply-chain breaches

Sources: Fireworks November 9, 2026 bulletin update · Fireworks October 9, 2026 bulletin · Google GTIG LLMjacking research