Morning Edition · №
Cybersecurity

A Second Breach at AI Infrastructure Startup Fireworks Spilled Into Notion's Chatbot

Attackers who compromised a third-party cloud provider used that foothold to reach Fireworks AI's development servers, and the fallout reached as far as real conversations inside Notion's AI assistant.

A Second Breach at AI Infrastructure Startup Fireworks Spilled Into Notion's Chatbot
Illustrative digital-code image; not a photo from the actual breach. — Photograph: Markus Spiske / Unsplash
SHARE X f in ⧉

Fireworks AI, a startup that rents out computing power to run open-source artificial intelligence models for companies including Notion, Cursor, Upwork and Vercel, has disclosed its second security incident in little more than a month — and this time the fallout reached into real conversations inside one of its customers' own products.

In an updated bulletin posted to its security advisories page on Nov. 9, Fireworks said an intruder who first broke into a third-party cloud provider used that access to reach the machines Fireworks uses for research and development, then ran an unauthorized workload that pulled customer data out the other side.

A Foothold Through a Cloud Provider

The intrusion traces back to Oct. 6, when Fireworks detected unauthorized access to environment variables — the API keys and tokens stored as secrets for evaluation jobs — tied to a limited number of customer accounts. The company disclosed that first incident on Oct. 9 and revoked the exposed credentials. Seventeen days later, while still investigating, Fireworks found a second, separate point of entry: a cloud provider hosting its R&D machines had suffered its own breach on Oct. 23, and the following day an unauthorized party used that foothold to access Fireworks' development systems directly.

The November update widened what Fireworks says was exposed: resource metadata, billing and shipping addresses, invoice and payment histories, account balances and credit-usage figures, plus — for an unspecified "limited set" of customers — incomplete subsets of AI model weights and training datasets. Fireworks says it does not store payment card numbers, so those were not affected, and that as of its most recent check it has found no further evidence of continued unauthorized access to the platform customers actually use. CEO Lin Qiao has said the attackers do not appear to have been after the model weights themselves — "most downloads were partial and stopped," according to coverage of the company's internal briefing — and that Fireworks has engaged Palo Alto Networks' Unit 42 to help investigate and lock down its systems. The company has attributed the activity to a financially motivated criminal group it says is interested in "LLM-jacking" — hijacking stolen API access to run up someone else's AI inference bill rather than steal the underlying technology.

The Notion Wrinkle

Where the breach stopped being an infrastructure story and became a privacy one was at Notion. In its own advisory, Notion said Fireworks notified it on Nov. 6 that the incident had affected fewer than 1% of Notion workspaces — specifically, transcripts of conversations run through Notion's AI "Insights" feature, which used Fireworks to serve open-weight models. Notion said most of the exposed transcripts carried no label identifying the author or originating workspace, but the content of the conversations themselves was readable. "This is not a breach of Notion's systems," the company's advisory states, framing the exposure as a vendor-side failure rather than an intrusion into its own infrastructure. Notion has since disabled Insights entirely, stopped routing open-weight-model traffic through Fireworks, and told customers to rotate any credentials or secrets they may have pasted into an AI chat.

At least one affected customer went public on their own: Render chief executive Anurag Goel posted on X that his company had received a notice confirming it was among those caught up in the exposure. The episode has drawn attention to how deeply embedded infrastructure vendors like Fireworks have become in products that look, from the outside, nothing like AI companies — a breach several steps removed from the user can still surface as their actual words on a screen.

The timing has not helped Fireworks' case that this is under control: it is the company's second disclosed breach since early October, and comes as a report from risk-rating firm Black Kite found that 34.2% of the more than 18,000 AI suppliers it tracks carry critical, unpatched vulnerabilities. For now, Fireworks says its investigation is ongoing and that it is working through notifications to every affected account; Notion and other customers are left doing their own audits of what, exactly, passed through a vendor most of their own users have never heard of.

SHARE THIS ARTICLE X Facebook LinkedIn Copy link
Claire Fontaine · Technology & Regulation Correspondent

Reports on technology and its regulation for UBStandard, with a focus on Brussels, AI policy and Europe's digital economy.

[email protected]
Related coverage Front page →