What happened
Fireworks’ November 17, 2026 security-bulletin update says the activity described in its November 9 and October 10 bulletins relates to two separate incidents. The company says both incidents are contained and remediated. It distinguishes them by timeline, route into its systems, customer impact and forensic status.
The first incident ran from August 3 through October 7. According to Fireworks, an unauthorized party probed its systems and eventually gained access. The party used stolen Fireworks credentials and a vulnerability in one of the company’s services to misuse the platform. For a small number of customers, the party accessed secrets, model files and datasets. Fireworks says Palo Alto Networks’ Unit 42 completed its forensic investigation of this incident. The November 17 update does not quantify the affected customer set or specify the vulnerability.
The second incident ran from October 23 through October 30. Fireworks says an unauthorized party gained access to its development machines hosted on a third-party cloud provider that had experienced a security incident. From those machines, the party reached part of Fireworks’ production systems. Account and billing information was accessed for all accounts created before October 30. For a small number of customers, datasets, models and/or secrets were also accessed. Fireworks says billing information did not include payment card numbers, which it does not store. AprosSyber is conducting the forensic investigation of this second incident; the update does not say that investigation is complete.
The security-bulletins page also describes earlier response measures and customer communications. In its November 9 update, Fireworks said it had notified law enforcement and engaged Unit 42, and described revoking and replacing credentials and service accounts, restricting network pathways, isolating affected systems, tightening data-storage access and enhancing logging and monitoring. It said it was notifying customers whose data was impacted and providing account-specific details and recommended actions. That earlier bulletin said self-serve dataset uploads and downloads for Batch Inference and dataset-based training had been disabled during the investigation. The November 17 update does not provide a restoration date for those functions.
What the archive adds
New Runtime’s earlier coverage reported that Fireworks’ October 2026 breach affected account and billing information, plus secrets, datasets and incomplete models for a limited set of accounts (November 10 note). A later note reported that attackers took Notion AI Insights transcripts from fewer than 1% of workspaces (November 13 note). The November 17 update is a new fact on that known story: Fireworks now says the activity comprised two distinct incidents and reports different investigative status for each.
Trend signal
This does not yet fit a pattern New Runtime tracks. The bulletin documents credential misuse and access through a third-party-hosted development environment, but it does not establish a broader change in how AI platforms manage agent identity, runtime boundaries or customer data. Treating one provider’s incident as proof of a general industry trend would go beyond what Fireworks has disclosed.
Connections to other events
The new update sits alongside New Runtime’s earlier reporting that Fireworks’ October incident exposed account and billing information, and that a limited set of customers also faced exposure of secrets, datasets and incomplete models. It also intersects with the reported Notion AI Insights transcript impact: New Runtime’s November 13 coverage said attackers took transcripts from fewer than 1% of workspaces. Fireworks’ November 17 bulletin does not name Notion or those transcripts, so it does not independently explain or revise that specific impact.
Why it matters
For teams using a hosted model platform, these disclosures show why an incident’s headline status is not enough to determine operational exposure. The first incident involved stolen Fireworks credentials and a service vulnerability; the second crossed from a third-party cloud provider’s environment into development machines and part of production. The customer data at issue also differs: account and billing information for all accounts created before October 30 in the second incident, and more sensitive datasets, models or secrets for a small number of customers across the incidents.
Customers should rely on account-specific notices for whether their own data or credentials were involved. Fireworks’ earlier bulletin advised affected customers to rotate identified credentials with the providers that issued them, update dependent applications and review activity accessible through those credentials. That matters because deleting a stored secret alone does not invalidate a credential at its issuer. Teams should also consider whether temporary restrictions on dataset uploads and downloads affect their workflows; the update gives no date for their return.
What happens next
- Fireworks will publish another update on the second incident’s investigation or remediation status (3 months, 70%). Watch for: a revised security bulletin identifying whether AprosSyber’s forensic work is complete.
- Fireworks will clarify whether self-serve dataset uploads and downloads have resumed (3 months, 60%). Watch for: a bulletin or product notice confirming restoration of Batch Inference and dataset-based training uploads and downloads.
Limits and uncertainty
Fireworks’ update does not state how many customers were in the “small number” whose secrets, model files, datasets or models were accessed, nor does it itemize which category applied to each affected customer. It gives no technical details about the service vulnerability, the stolen credentials, or the third-party cloud provider. Unit 42’s investigation of the first incident is complete, but AprosSyber’s investigation of the second is ongoing. The update describes both incidents as contained and remediated, but does not publish independent confirmation of that status or quantify any data exfiltrated. Fireworks says it does not store payment card numbers; the update does not state whether other billing details were misused.
Skill to build now
If your team uses Fireworks, inventory the applications and jobs that depend on platform secrets, dataset uploads or dataset-based training. If you receive an incident notice, rotate each identified credential with its issuing provider, then update the application that uses it; deleting a copy from a platform does not revoke the credential itself. Review logs for activity available through the exposed credential, and check whether paused dataset workflows need a temporary alternative.
Reader takeaway
Fireworks’ update separates what had been described as October security activity into two incidents with different entry routes and investigation status. Both are reported as contained and remediated, but one exposed account and billing information across pre-October 30 accounts, while only a small number of customers faced access to datasets, models or secrets. Affected customers still need account-specific details to assess their own exposure.