THE ESSENTIALS IN 30 SECONDS
- An attacker used an incident at a third-party cloud provider hosting Fireworks' R&D machines to reach its development systems on 24 October; it was Fireworks' second incident since early October.
- Fireworks says billing data, resource metadata and, for some customers, datasets, incomplete model weights and secrets were accessed; it links the activity to a financially motivated LLMjacking group.
- Notion says AI chat transcripts tied to its Insights feature and a volunteer testing programme were exposed for fewer than 1% of workspaces, despite a contractual zero-retention commitment.
Fireworks AI, a platform that developers use to run open-weight models, has confirmed a second security incident in two months, and Notion has confirmed that some of its users' data was caught up in it. TechCrunch reported on 12 November that hackers obtained chat transcripts from some Notion AI users.
What Fireworks disclosed
In an updated security bulletin dated 9 November, Fireworks says a third-party cloud provider hosting its research and development machines suffered an incident on 23 October. The next day, an unauthorised party used that foothold to reach Fireworks development machines and run a workload of its own. Fireworks says it detected the activity the same day. The provider is not named.
The update widens the scope of an incident first disclosed in a 9 October bulletin, which reported that secrets used in evaluation jobs had been accessed. Fireworks now lists account and billing data (addresses, invoices, balances and credit usage) and resource metadata as affected, along with datasets, incomplete model weights and secrets for a limited set of accounts. It says it does not store card numbers and, so far, has seen no impact on its inference, Training SDK or Nexus products. Fireworks has brought in Palo Alto Networks' Unit 42 and informed law enforcement.
The company links the activity to a financially motivated group practising what it calls LLMjacking: obtaining access to language models and reselling the tokens to illicit routing services. BleepingComputer notes that Google's threat intelligence team and Mandiant had warned about such attacks a week earlier.
Notion's exposure
Notion says Fireworks told it the incident affected fewer than 1% of Notion workspaces. The exposed material consists of Notion AI chat transcripts, meaning model inputs and outputs written to the chat, from its Insights feature, which classifies AI usage for admins, and from volunteers in its Learning and Early Access Program. Meeting-note audio and transcripts were not affected unless pasted into a chat. Notion adds that Fireworks had contractually committed to zero data retention, yet data from the endpoints Notion used was kept.
On 3 November, Notion cut its connection to Fireworks, deleted its data there, halted training runs and disabled Insights. It has contacted affected customers and advises users to rotate any passwords or API keys pasted into Notion AI chats. According to TechCrunch, Notion says much of the data cannot be tied to a specific customer.
