Fireworks AI, a platform for running and training AI models, disclosed a second breach in about a month. This time the attackers came in through an outside cloud provider that hosted the company's research and development machines, and according to Notion, one of its customers, transcripts of conversations with Notion AI were exposed too.
What happened?
According to Fireworks, the cloud provider was compromised on October 23, and a day later the attackers used that access to reach the company's development systems and run an unauthorized workload on them. They accessed resource metadata, such as identifiers and names, and customer billing details, including addresses and payment history. For a limited number of accounts, secrets, datasets and incomplete models were also exposed. The company did not say how many customers were affected or which cloud provider was involved.
And what about Notion?
Notion said fewer than 1% of its workspaces may be affected. According to the company, the data is Notion AI chat transcripts used for a feature called Insights, not attributed to a user or workspace, and without page content or external data Notion AI pulled in. Notion stresses this is not a breach of its own systems. According to Bloomberg, Fireworks is also assessing the impact with other customers, including Cursor and Harvey.
Who is behind it?
Fireworks believes the attackers are linked to criminals engaged in LLMjacking: stealing access to AI models and selling it through illicit gateways. That is the same phenomenon behind the wave of breaches at AI platforms in October. The company hired Palo Alto Networks' Unit 42 to investigate.
The lesson
This story shows how risk moves along a chain: a Notion user never chose Fireworks or heard of it, but their conversations passed through it. Any company building AI into its product, including many Israeli companies, needs to ask not only where its data is stored, but which vendors it passes through and what they keep after the conversation ends.
