One incident becomes two
Fireworks AI first detected unauthorized use of internal credentials on October 6 and disclosed the incident on October 9, reporting that attackers had accessed environment variables in evaluation jobs, including API keys and tokens belonging to some customers. At that point, the company said it had no evidence of access to models, datasets, training data, inference traffic, or personal information, a disclosure that read as contained to a specific, lower-severity blast radius.
That containment did not hold. On October 23, a third-party cloud provider hosting Fireworks' research and development machines was itself compromised, and attackers used that access the following day to reach Fireworks development systems and run an unauthorized workload. Fireworks did not publish the broadened impact of this second compromise until November 9, meaning the full picture took more than two weeks to reach customers after the second intrusion actually occurred.
What the expanded disclosure actually covers
The November 9 update is substantially wider than the October 9 one. It covers resource metadata, including identifiers, display names, and creation timestamps, plus billing information: billing and shipping addresses, invoice and payment histories, account balances, and credit usage. Fireworks says payment card numbers were not stored and are unaffected, which limits the direct financial fraud risk even as it confirms a meaningful amount of account-level business data was exposed.
More significant for an AI infrastructure vendor, the update confirms that for certain customers, a subset of incomplete model weights and datasets was accessed. Fireworks CEO Lin Qiao has said attackers did not appear to be primarily after weights, noting that most weight downloads observed were partial and stopped before completion, a detail meant to reassure customers that no one walked away with a usable copy of their proprietary model.
The Notion connection nobody asked for
The most unusual element of this disclosure is Notion's involvement. Notion told at least one affected customer that some unattributable chat sessions with its own AI assistant may have been exposed through this incident, affecting under 1 percent of Notion workspaces. Notion says the exposed transcripts were not linked to specific users or workspaces, but users on Reddit have pointed out that chat content itself can still contain identifying details even without an explicit account link attached.
This is the clearest illustration in the incident of how deeply AI infrastructure vendors are now entangled with each other. A breach at an inference provider like Fireworks rippled into a completely separate product, Notion AI, because Notion runs some of its AI features on Fireworks infrastructure. That is the structural risk every enterprise adopting AI products now carries: your vendor's vendor is part of your attack surface, whether or not it appears anywhere in your own vendor risk register.
A financially motivated, not espionage-driven, actor
Fireworks suspects the attackers are financially motivated criminals involved in LLMjacking, the practice of gaining unauthorized access to AI models and reselling that access through illicit API gateways to other paying users. That motive profile is consistent with the partial, interrupted weight downloads Lin Qiao described: an attacker optimizing for resale value and usage arbitrage behaves differently than one trying to exfiltrate a complete proprietary model for competitive or state purposes.
Fireworks worked with Palo Alto Networks' Unit 42 on remediation, rotating credentials, isolating affected systems, and tightening data storage access and monitoring. Bloomberg has reported that Fireworks is working with customers including Cursor and Harvey to assess impact, naming two companies whose own products sit directly in front of enterprise users and legal professionals respectively.
What this means for every company building on inference APIs
Model weights joining billing data and API keys as a routinely disclosed breach category is a shift worth noting on its own. Enterprises that fine-tune or host proprietary models through a third-party inference provider now need to ask that provider directly how weights are stored, whether they are encrypted at rest separately from the serving infrastructure, and what access a compromised build or research environment would actually have to production weight stores.
The Notion angle adds a second question that is harder to answer from a vendor questionnaire alone: which of your AI vendors run on top of which other AI vendors, and does a breach disclosure from one of them obligate the other to tell you anything. Right now, that obligation appears to run through goodwill and reputational pressure rather than contract, which is not a comfortable place for any enterprise relying on Fireworks and Notion between them.
A disclosure timeline worth scrutinizing
Fireworks detected unauthorized credential use on October 6, disclosed on October 9, suffered a second, broader compromise through its cloud provider on October 23 and 24, and did not publish the widened scope until November 9, more than two weeks after the second intrusion. That gap is not unusual by industry norms, forensic investigation of a multi-stage breach takes time, but it means customers spent over two weeks operating on an initial disclosure that understated the actual exposure.
Any enterprise that received the October 9 notice and concluded its exposure was limited to environment variables and API keys needs to revisit that assessment now that model weights, datasets, and billing records are confirmed in scope. The practical takeaway is to treat an initial breach disclosure from any vendor as a floor, not a ceiling, and to ask explicitly whether the investigation is ongoing and the scope could still widen, rather than filing the first notice away as the final word. That question costs nothing to ask and would have given Fireworks customers a two-week head start on the remediation they are only now able to begin in full.



