October Incident Update
Here's my update to the broader community about the October security incident. I want to give you visibility into the situation directly. We also published a detailed update to the security bulletin: https://fireworks.ai/security/bulletins/october-2026-security-incident-update
A third-party cloud provider that hosted Fireworks development machines experienced a security incident on October 23. On October 24, an unauthorized third party used this access to reach a Fireworks development machine. From there, they escalated permissions via multiple hops to assume one of the cloud service accounts. This allowed them to access a subset of storage systems, leading to additional customer data impact. We detected this activity on the same day and followed our incident response processes to take immediate investigation and remediation measures. As of October 25, we have not detected further evidence of successful unauthorized access by the threat actor on the Fireworks platform customers use.
In response to the incident, we notified law enforcement and obtained assistance from a leading cybersecurity firm, Palo Alto Networks Unit 42, to help conduct forensic analysis. When we published the first bulletin, only credentials stored in Fireworks (e.g., those used in evaluation jobs) were known to have been accessed. Since then, we have discovered impacts to account metadata, and billing information (payment card numbers are not impacted.). Our investigation also found that other customer data was exposed in connection with the initial incident, affecting a small number of accounts. Specifically, for a limited** **set of accounts, we discovered impacts to credentials, datasets, and models (most access is incomplete). We notified impacted customers with information on the range of impact, specific details, and required follow-up actions.
While we investigate this incident, we’ve disabled self-serve dataset uploads and downloads used by batch Inference and dataset-upload-based training. We have not detected any impact to Inference (excluding batch inference), Training SDK, and Nexus, which all have zero data retention (ZDR) by default. Training SDK remains available and intact with ZDR, i.e. no training data stored at rest, and Fireworks Lab-based training operates in an isolated environment with additional network and storage protection using Palo Alto Network firewalls. We continue to work with Palo Alto Networks Unit 42 on investigation as needed.
The evidence indicates links to a financially motivated cybercriminal threat actor group that aims to sell tokens from unauthorized access to Large Language Models (LLMs) to illicit routing gateways, known as "LLMJacking". On November 2, Google Threat Intelligence Group and Mandiant published a blog post about the rise of LLMJacking, their motives, techniques used, and mitigation. Based on the evidence we see, we believe the attacking group is sophisticated in using AI to drive the attack at high speed.
AI is fundamentally changing the cyber landscape. We saw a big part of the defense system shift toward stronger collaboration across the industry and building a community-wide immune system. As we researched the industry state of the art, we identified patterns of compromise and vulnerabilities in several other companies' infrastructure, for example, exposed internal credentials in their public Docker images. In those cases, we shared their vulnerability with the affected companies immediately. In our incident response, we used AI to be a detection and defense tool, able to identify patterns across multiple systems and craft dedicated detectors for suspicious activity. Agents are capable pen-testers and can help you find extremely complex vulnerabilities before attackers can. We’ve been using several products powered by closed frontier (such as OpenAI Daybreak, CrowdStrike Falcon Sensor and Wiz Sensor) and open-weight models to do red teaming and offense detection. Here are additional practices we strongly encourage adoption: 1) Long-lived credentials have long been considered bad security practice, as one leak compromises the entire identity. But, they proliferated with modern cloud-based platforms, including AI companies. Workload identity federation with short-lived, cryptographically signed tokens is a strong alternative, as is frequent key auto-rotation. Isolating networks with private links or IP-level filtering adds another layer of protection. 2) Systemic enforcement of best security practices is crucial given the faster pace of agent-assisted offensive hacking, including aggressive, regular and proactive rotation of credentials, regular audits of production access controls, fine-grained identities and permission sets for internal infrastructure, strict network access policies, strict tenant isolation, etc. Programmatically enforcing security policy, running regular automated audits, and using agents for scans are must-have.
I thank companies jumping in to support us, including OpenAI, Vercel, Palo Alto Networks, CrowdStrike and Wiz to access the best tools and share knowledge, and many more companies reaching out to help. Thank you for the team work!


