Hackers Raided An AI Platform Trusted By 10,000 Companies. Is Yours Next?

Today, we're diving into:

Hackers Raided An AI Platform Trusted By 10,000 Companies. Is Yours Next?

Fireworks AI, which hosts and fine-tunes open models for companies including Uber, Shopify and DoorDash, says hackers accessed some customers' datasets and models, Forbes reported on November 9. Forbes puts its customer base at about 10,000 enterprises.

The incident began in early October. Fireworks first said only a small set of secrets had been touched. Its updated security bulletin, dated November 9, now describes a much wider picture.

You Cannot Protect Data You Never Knew You Shared.

Find out which datasets, models and keys your AI vendors are holding right now.

What Fireworks Says Happened

According to the original bulletin, Fireworks spotted unauthorized use of internal credentials on October 6 and started notifying customers the next morning. At that point, it said, only secrets stored in evaluation jobs for a limited number of accounts were affected.

The story did not end there. Fireworks says a third-party cloud provider hosting its research machines had its own security incident on October 23. A day later, the attacker used that foothold to reach Fireworks development machines and run a workload that exposed more customer data.

Fireworks links the attack to a financially motivated group practicing "LLMjacking": stealing access to AI models and reselling the tokens through illicit gateways. Forbes reports that CEO Lin Qiao said on X the attack appeared to involve AI agents.

Customers Are Already Voting

Two customers told Forbes they had moved a large share of their work off the platform. Others said they had heard nothing about the incidents at all.

This Is Not A Vendor Problem. It Is A Data Inventory Problem.

The easy reaction is to blame one startup and switch providers. That misses the lesson.

Every time your team fine-tunes a model, runs an evaluation or uploads a test set, a copy of your data and your secrets lands on someone else's infrastructure. Most enterprises cannot say which datasets went where, who approved it, or which keys were baked into the jobs.

The real bottleneck is not your vendor's security team. It is your own record of what you handed over.

The wrong reaction is to freeze all external AI work. Your engineers will simply find another platform with less scrutiny.

The better alternative is to treat every AI vendor as part of your data estate, with the same inventory, ownership and lineage you demand of your own warehouse.

What You Should Do This Week

Knowing what you shared starts with trusted enterprise context. DataManagement.AI helps organizations create that foundation by unifying metadata, lineage, governance, business glossaries and knowledge assets into a single intelligence layer. It automatically discovers metadata, maps lineage across the modern data stack and identifies downstream dependencies, so when a vendor reports a breach, you can see in minutes which of your datasets were exposed and what depends on them.

Google Just Built An AI Chip For Your Most Private Data. Do You Need It?

Google has added two chips to its eighth-generation TPU lineup, The Next Web reported on November 9. One is built for secrecy. The other is built for size.

That brings the eighth generation to four systems, after the TPU 8t and 8i were unveiled at Cloud Next in April. Google has not disclosed pricing or release dates for either new chip.

What Google Announced

The TPU 8tx is aimed at governments and regulated businesses that want to run models and agents on private data. Security is built into the hardware, keeping data and model weights isolated while they run. It works alongside AMD's fifth-generation EPYC processors, and Google says its network design cuts latency by up to 50% compared with the TPU 8t.

The TPU 8i Max puts two TPU 8i chips in one package, doubling compute, on-chip memory and memory bandwidth. Google is pitching it at agents that work across entire codebases and systems that analyze hours of video.

Confidential Computing Is Moving Into The AI Stack

Until now, most "private AI" promises were contractual. You trusted the provider's policy that your data would not be seen or kept.

This fundamentally changes the conversation. If isolation is enforced by the chip, the question shifts from "do you promise?" to "can you prove it?", which is exactly what your risk and compliance teams have been asking.

The Real Question Is Which Workloads Deserve It

The wrong reaction is to move everything onto confidential hardware because it sounds safer. Specialized chips usually come at a premium, and Google has not said what that premium will be.

The better alternative is to sort your AI work by sensitivity first. Fine-tuning on patient records or trading data may justify sealed hardware. Summarizing public product manuals does not.

Know Your Inference (KYI) is built for exactly this decision. When you measure how accurate, fast and costly each AI workload is on your own cases, you can price the cost of confidentiality per job, instead of paying for it everywhere.

ChatGPT Enterprise Just Landed On Your AWS Bill. Who Approved It?

OpenAI has made ChatGPT Enterprise available in AWS Marketplace, PPC Land reported on November 9. Companies can now buy it through their existing AWS account.

The listing covers the same models and features as buying direct, including ChatGPT and Codex, with new features arriving at the same time. Eligible customers can also get ChatGPT for Healthcare and ChatGPT for Financial Services.

Why This Is A Bigger Deal Than It Sounds

According to PPC Land, the listing says eligible spend counts toward a customer's AWS commitment, subject to their AWS agreement. Pricing is custom and handled through private offers.

For many enterprises, that removes the hardest part of buying AI: a new vendor review, a new contract and a new budget line. Money already committed to AWS can now flow to OpenAI.

Faster Buying Is Not The Same As Better Governance

Procurement friction was annoying. It was also one of the few places where someone asked who would use the tool, what data it would see and who owned the risk.

PPC Land notes the listing says business data is not used to train OpenAI's models by default, and that OpenAI acts as the data processor, using subprocessors including AWS. Those are good defaults. They do not tell you which of your teams just turned it on.

The Wrong Reaction Is Treating It As Just Another SaaS Line

When AI becomes as easy to buy as storage, it spreads like storage. A few teams start, usage grows, and a year later nobody can explain what the spend is buying.

The better alternative is to set the rules before the first private offer is signed.

The real bottleneck is no longer getting AI approved. It is knowing what you are getting back once it is everywhere.

Your opinion matters!

Hope you loved reading our piece of newsletter as much as we had fun writing it.

Share your experience and feedback with us below 'cause we take your critique very critically.

Thank you for reading

-Shen & Towards AGI team