Why do customers stay with a vendor that has now been breached twice in five weeks? Because for many of them, leaving costs more than the leak.
On October 24, an intruder got into Fireworks AI's development machines through a third-party cloud provider. The company's own security bulletin, sent to customers on Monday, November 9, says the attacker reached billing records, account details and incomplete model weights. Bloomberg reported that it is the Nvidia-backed company's second known breach in weeks, and that Fireworks tied the attack to a financially motivated group that steals access to AI models and resells it, a practice known as LLMjacking.
The damage, on the company's account, is bounded. Payment card numbers aren't stored by Fireworks and weren't touched. Customer-facing services kept running, and the company says inference, its Training SDK and Nexus, which keep no customer data by default, showed no impact. Fireworks says it detected the activity the same day, replaced credentials and service accounts, isolated affected systems, and notified law enforcement and customers. The exposed set also includes resource metadata and, for a limited group of customers, datasets.
That is a competent response. It is also the second one in a month.
On October 9, Fireworks had already disclosed that someone used internal credentials to read environment variables in a limited number of customer accounts, and that API keys and tokens stored as secrets were probably captured. The October 24 intrusion came 15 days after that bulletin. Runtime Wire called it the company's second security breach in weeks, and PANews, citing Bloomberg, reported that hackers had been inside customer datasets and models over the past month.
Meta strategy chief Dave Wehner leaves to become CFO of Fireworks AI Fireworks AI, which serves models for Cursor, Perplexity and Uber, hired Dave Wehner after he spent a decade as Meta's CFO and then strategy chief. The startup was valued at $17.5 billion in July. Neither side has given a start date. - Why did Dave Wehner leave Meta - Fireworks AI new CFO Dave Wehner
A vendor that holds your unfinished model weights isn't holding a login. It's holding months of GPU time and training decisions. Incomplete weights are exactly what an LLMjacking crew or a competitor would want, and unlike a credit card, you can't cancel them and get a new set.
Why the customers stay #
Fireworks says it has over 10,000 customers. A Fireworks spokesperson told Bloomberg the company is working with clients on the incident, naming Notion and Cursor, and PANews' summary of the Bloomberg report lists Uber and Shopify among its users. Fireworks was valued at $17.5 billion as of July. Nothing in this breach has been reported to dent that.
No customer exodus has been reported. The reason is the market, not a lack of anger. Fireworks serves open models at speeds and prices its customers have tuned their products around, and the capacity to run that traffic elsewhere is scarce. A company like Cursor can't switch providers on a weekend because a bulletin landed. It would have to find equivalent GPUs, re-benchmark latency, move its fine-tuned weights and keep serving users the whole time, in a market where chips are tight.
Frankly, that is the real exposure. Security incidents usually cost a vendor customers because switching is cheap enough to punish bad behaviour. Here it isn't. The customers with spare options can leave, and the ones who can't are the ones with the most weights sitting on Fireworks machines.
The pressure Fireworks feels is reputational and contractual, and that's softer than churn. Notion and Cursor will ask for audits, shorter data retention and tighter access terms, and some will get them. The forensic findings may produce a harder look at which research and development systems ever touch customer material. Those are real costs, but they are small ones next to what a reliable exit would impose.
If you run a product on a hosted inference provider, treat that as a planning fact. Keep copies of your weights and datasets under your own keys, ask what the vendor's R&D environment can reach, and price a second provider before you need one. The incident itself is closed on Fireworks' side, and no group has been named beyond the LLMjacking description. Also read: Meta Steals the Spotlight From Anthropic's Fable 5.2 With Muse Ember, Its Smallest and Cheapest AI Model • CoreWeave faces earnings as neocloud debt, customers and politics all bite • Akamai launches $2.2 billion convertible bond sale to repay 2027 debt and fund cloud build
Notion agents can now run unsupervised for 24 hours, up from five minutes Notion credits a dedicated, queue-backed worker service for the leap, and says its upgraded Auto router now matches Opus 5.5 and Astra at roughly 90% lower cost across 20-plus common tasks. - Notion AI agents run 24 hours unsupervised - how long do Notion agents run autonomously
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.