Fireworks AI Hit by Another Cyberattack, Notion User Chat Transcripts Leaked

Fireworks AI hit again within two months; Notion customer chat transcripts accessed without authorization.
AI infrastructure provider Fireworks AI has suffered another cyberattack, allowing attackers to access Notion customers' chat transcripts—its second breach in two months. Fireworks occupies the middle layer of the AI application stack, serving apps like Notion and orchestrating GPU compute and models, so it holds sensitive conversation data and compute resources. Chat transcripts contain natural language input that may involve trade secrets and privacy, and leaks propagate down the supply chain to downstream customers and end users. The incident exposes security governance gaps among third-party providers and signals that enterprises should weigh security records and incident response when choosing inference vendors.
AI infrastructure provider Fireworks AI has once again been targeted by a cyberattack. According to disclosures, the breach allowed attackers to access Notion customers' chat transcripts. This is the second data breach at Fireworks in just two months, and the successive security incidents have plunged the platform—which provides model inference services for AI applications—into a crisis of trust.
Incident Recap: Two Breaches in Two Months
Based on available information, the core risk in this attack is that Notion customers' chat transcripts processed through Fireworks AI were accessed without authorization. Notion, a widely used collaboration and knowledge management tool, relies on third-party inference services behind its AI features, and Fireworks is one of the providers of such services. When the underlying infrastructure is compromised, user data in the applications above it is exposed to risk.
You may not have noticed that this is not an isolated incident. The source material explicitly states that this is Fireworks' "second data breach in two months." Consecutive security incidents suggest the problem may not be an occasional oversight but rather a systemic weakness in security governance.
Why AI Inference Platforms Become Attack Targets
Platforms like Fireworks AI sit in the middle layer of the AI application stack: upward, they serve the product needs of applications like Notion; downward, they orchestrate GPU compute and models. This position means they hold two types of high-value assets at once—customers' prompts and conversation content, as well as model invocation and compute resources.
Chat transcripts are especially sensitive. Unlike ordinary logs, chat transcripts often contain users' natural language input, which may involve trade secrets, personal privacy, or even identity information. Once leaked, the impact is not limited to the platform itself but extends to downstream customers like Notion and their end users.
Model inference refers to the process of deploying a trained AI model into a production environment, performing real-time computation on user input, and returning results. Unlike the model training phase, which relies on large-scale GPU clusters and takes weeks, inference services emphasize low latency, high concurrency, and elastic scaling, and are typically offered externally as APIs. Vendors such as Fireworks AI reduce per-call costs by optimizing inference engines (e.g., operator fusion, dynamic batching, quantization compression), enabling application developers to invoke large models without building their own compute infrastructure. Precisely because inference services directly handle raw user input in the request chain—including prompts, conversation context, and file content—they are naturally a high-risk node for data breaches. Once attackers compromise an inference platform, they may obtain conversation data across customers in bulk, with an impact far greater than a breach of a single application.
Knock-On Effects on Downstream Customers and Users
A key feature of this incident is the transmission of responsibility along the chain. Notion's users may never have dealt directly with Fireworks, yet their data flows through third-party services. This exposes a visibility problem common in the AI supply chain: users at the application layer can hardly know which links their data actually passes through.
For enterprises that rely on third-party inference services, this incident raises a practical question—when choosing AI infrastructure vendors, security records and incident response capabilities should be included as core evaluation dimensions, not just price and performance.
AI supply chain security refers to the security risks at each link of the dependency chain that makes up an AI application—from underlying compute, model hosting, and inference services to the products above—and how those risks propagate. Traditional software supply chains focus on vulnerabilities in open-source components and poisoned dependency packages, whereas AI supply chains additionally involve model weight protection, training data provenance, and the flow of sensitive information in inference requests. Because AI applications generally adopt a layered architecture of "application layer—inference layer—compute layer," end users often interact only with the topmost product and lack visibility into the third-party services their data passes through. This means that once a middle-layer vendor is compromised, downstream customers can neither detect it in time nor audit it independently. At the regulatory level, the EU AI Act and data protection laws in various countries are gradually requiring enterprises to disclose third-party participants in their data processing chains, and supply chain transparency is shifting from a best practice to a compliance requirement.
Consecutive Breaches Sound the Alarm
Two breaches in two months deal a direct blow to Fireworks' brand and customer trust. In an increasingly competitive AI infrastructure market, being secure and trustworthy is becoming a key differentiator. For the industry as a whole, this incident also serves as a reminder: as AI capabilities are embedded in more and more productivity tools, the security level of middle-layer vendors will directly determine the baseline of data security for end users.
Public information is currently limited, and further disclosure is still awaited regarding the specific attack methods, the number of affected users, and Fireworks' response measures.
One-line Summary
Fireworks AI hit by a second attack within two months, with Notion customer chat transcripts accessed without authorization.
Paragraph Summary
AI infrastructure provider Fireworks AI has been hit by another cyberattack, this time allowing attackers to access Notion customers' chat transcripts—the company's second data breach in two months. Fireworks sits in the middle layer of the AI application stack, serving applications like Notion upward and orchestrating GPU compute and models downward, and thus holds both sensitive conversation data and compute resources. Chat transcripts contain users' natural language input and may involve trade secrets and personal privacy, with the impact of a leak propagating along the supply chain to downstream customers and end users. The incident exposes shortcomings in the security governance of third-party vendors in the AI supply chain and suggests that enterprises should include security records and incident response capabilities as core evaluation dimensions when selecting inference service providers.
Related articles

NVIDIA's Two Faces: Strong Earnings and Executives Calling for Jensen Huang's Return
NVIDIA's latest earnings are strong, but tech executives are publicly calling for Jensen Huang's return. This article analyzes the industry anxiety and ecosystem tensions behind these two narratives surrounding the AI chip giant.

Oral Pancreatic Cancer Drug Hailed as Biggest Breakthrough in Decades
An oral pancreatic cancer drug is called the biggest breakthrough in decades, sparking Hacker News debate. We analyze the real weight of this claim and the distance to clinical availability.

DevWorkflow 5.0: From Idea to Launch with the AI Five-Step Method
DevWorkflow 5.0 is a Claude Code plugin that uses a five-step process to take AI-assisted development from idea to launch. It asks you before every decision, preventing AI from adding tables and fields on its own, keeping docs and code consistent, and stopping merges before tests are written. Supports multiple platforms and languages, for individual developers and small teams.