OTWopentechwire
Tech Intelligence, Openly Wired
Startups

Supply Chain Breach at Fireworks AI Exposes Notion Chat Data

A November intrusion at the AI infrastructure provider compromised customer billing records and datasets, highlighting the cascading security risks in developer platforms that serve thousands of downstream users.

HP
Hana Park
Semiconductors Reporter · Seoul
Nov 16, 2026
7 min read
Supply Chain Breach at Fireworks AI Exposes Notion Chat Data
Credit: TechCrunch

Infrastructure Provider Compromised Through Cloud Layer

An intrusion at Fireworks AI, a platform that brokers access to open-source large language models, has exposed customer information including billing details and training datasets, the company disclosed in a security advisory this week. Attackers penetrated the company's cloud infrastructure to reach a development environment, from which they extracted data belonging to multiple clients.

Fireworks AI operates as middleware in the generative AI stack, offering developers streamlined access to open models without maintaining their own inference infrastructure. The company counts tens of thousands of customers globally, among them legal AI firm Harvey and coding assistant Cursor. The breach demonstrates how a single compromise at the infrastructure layer can ripple outward to affect end-users several steps removed from the initial point of failure.

The intrusion occurred in early November. Fireworks has not specified which cloud provider was breached, nor has it detailed the method attackers used to move laterally from cloud infrastructure into its own development systems. The company's advisory confirmed that both billing information and certain datasets were taken, but it provided no estimate of the total volume or the number of customers affected.

Notion Confirms Customer Transcript Theft

Among the affected downstream customers, productivity software maker Notion has publicly acknowledged that roughly one per cent of its user base was impacted. Fireworks alerted Notion to the incident, prompting the company to sever all connectivity with the AI provider on 3 November, delete data stored in Fireworks systems, and halt integration with the service.

Notion confirmed this week that stolen data included chat transcripts generated by customers using its AI-powered Insights feature. Those transcripts capture both prompts submitted to the model and responses returned, creating a complete record of user interactions with the AI assistant. Notion noted that much of the exfiltrated data could not be definitively linked to specific customers, complicating notification and remediation efforts.

The Insights feature, which helps users query and synthesise information across their Notion workspaces, relies on external model inference provided by Fireworks. This architectural choice, common among software vendors seeking to add AI capabilities without building in-house model serving, creates a dependency that extends the attack surface beyond the company's direct control.

Token Resale as the Likely Motive

Fireworks AI stated in its advisory that the attackers appear to be financially motivated, with their primary objective centred on obtaining tokens that grant access to large language models. These tokens, once stolen, are sold to unauthorised third parties who can then consume model inference capacity without incurring costs themselves. The scheme effectively allows buyers to run AI workloads on someone else's compute budget, a form of resource theft that has grown alongside the proliferation of model-as-a-service platforms.

This motive distinguishes the breach from data exfiltration campaigns aimed at intellectual property or customer espionage. The attackers' focus on monetising access rather than hoarding proprietary information suggests they view AI infrastructure as a commodity to be arbitraged. As model usage costs remain high, particularly for frontier models and high-throughput applications, the black market for stolen inference capacity has become more attractive to cybercriminals.

The October breach at Fireworks, which the company has not detailed publicly, indicates a pattern of persistent targeting. Two intrusions within a span of weeks may reflect either inadequate remediation after the first incident or a broader campaign by threat actors who have identified Fireworks as a high-value target within the AI supply chain.

Cascading Risk in the AI Stack

The incident illustrates a structural vulnerability in the contemporary AI ecosystem: most applications do not host their own models but instead route requests through intermediaries like Fireworks, which in turn rely on cloud providers for compute and storage. Each layer introduces potential points of compromise, and a breach at any one of them can propagate downward to affect end-users who have no direct relationship with the breached entity.

For Notion's customers, the path from their chat input to a stored transcript involved at least three parties: Notion itself, Fireworks AI, and Fireworks' unnamed cloud provider. The attacker needed to compromise only one link in that chain to reach data generated by users of a consumer productivity tool. This architecture is efficient and cost-effective, but it distributes security responsibility across multiple organisations with varying levels of transparency and incident response capability.

Enterprises evaluating AI integrations increasingly face a dilemma: build inference capacity in-house, which requires significant capital and engineering resources, or rely on third-party providers and accept the attendant supply-chain risk. For most startups and mid-market companies, the former option remains prohibitively expensive, leaving them dependent on platforms like Fireworks, OpenAI's API, or hyperscaler model services.

Limited Disclosure and Customer Notification Challenges

Neither Fireworks nor Notion has published a detailed timeline of the breach, including when the intrusion began, how long it persisted before detection, or what specific controls failed. Fireworks' advisory mentions only that customers were notified and that the company has taken steps to secure its systems, without elaborating on those measures.

Notion's statement that much of the stolen data is unattributable complicates the company's legal and ethical obligations. Privacy regulations in multiple jurisdictions require notification when personal data is compromised, but if the company cannot determine which users were affected, it must either notify all potential victims or rely on probabilistic assessments. The one per cent figure suggests Notion has some ability to scope the impact, but it is unclear whether that estimate is based on server logs, data volume analysis, or another method.

The lack of detail about the cloud provider involved is notable. If the initial compromise occurred at a major hyperscaler, the incident could have broader implications for other customers of that provider. If it was a smaller or regional cloud vendor, the breach may reflect less mature security practices. Fireworks' decision not to name the provider leaves open questions about whether the cloud vendor has acknowledged the breach or taken independent action.

Broader Implications for Model-as-a-Service Security

The Fireworks breach arrives as the model-as-a-service market continues to expand, with dozens of platforms competing to offer developers easier, cheaper access to open and proprietary models. Many of these providers operate with lean security teams and prioritise speed to market over hardened infrastructure, a trade-off that becomes more consequential as customer counts grow.

Stolen tokens and resold inference capacity represent a new category of cybercrime that traditional security frameworks are not well equipped to address. Unlike credit card fraud, where financial institutions can freeze accounts and reverse transactions, stolen API tokens can be used immediately and at scale, with limited recourse for the victim. Model providers can revoke compromised tokens, but doing so disrupts legitimate customers and requires rapid detection, which many platforms lack.

The incident also raises questions about data retention policies at AI intermediaries. Fireworks stored customer datasets and chat transcripts in its development environment, a practice that may be necessary for debugging and model fine-tuning but also increases the volume of sensitive information at risk in a breach. Minimising data retention and segregating production from development environments are standard security practices, but the advisory does not clarify whether those controls were in place or bypassed.

What Happens Next

Fireworks has not disclosed whether it has engaged external forensic investigators or shared indicators of compromise with industry peers. Notion has moved to a different AI provider or brought inference in-house, though the company has not specified its new architecture. For affected Notion customers, the company has advised reviewing chat transcripts and assessing whether any sensitive information was included in prompts or model outputs.

The breach is likely to accelerate scrutiny of vendor security postures among enterprises adopting generative AI tools. Procurement teams are beginning to ask for third-party audits, penetration test results, and attestations of compliance with standards such as SOC 2 or ISO 27001. Platforms that cannot demonstrate mature security practices may find themselves excluded from enterprise deals, even if their models or pricing are competitive.

Regulatory attention may follow. Data protection authorities in the European Union and United Kingdom have indicated that AI supply-chain incidents fall within the scope of GDPR and the UK Data Protection Act, and that both direct processors and subprocessors can be held liable. If affected users include EU or UK residents, Notion and Fireworks may face inquiries from supervisory authorities.

For now, the incident serves as a reminder that the AI stack is as vulnerable as its weakest component, and that security failures at infrastructure providers can cascade into customer-facing applications with little warning. As the ecosystem matures, the industry will need to develop better mechanisms for auditing third-party providers, isolating sensitive workloads, and responding to breaches that span multiple organisations.

Read next
Startups

Nvidia's Open Source Strategy Faces Uncertain Future After Huang Resignation

Arjun S. Mehta · 7 min
Startups

Colette Kress Takes Helm at NVIDIA as Huang Resigns Amid Probe

Hana Park · 4 min
Startups

Colette Kress Steps into Nvidia's Top Job After Huang's Sudden Exit

Arjun S. Mehta · 4 min
Spot something wrong? Email corrections@opentechwire.com. We log every correction publicly.