← Back to blog
GeneralAbout 6 min read

Black October: Nine AI Platforms Lost Keys, Not Code

Published Nov 10, 2026
Black October: Nine AI Platforms Lost Keys, Not Code

Nine AI developer platforms disclosed security incidents in October. Wiz Research started calling it Black October in a post published October 30, and the name stuck because the pattern was so consistent it looked coordinated even though nobody has established that it was.

The list, in disclosure order: Fireworks AI on October 6, Andromeda on the 9th, Zeabur on the 14th, Lightning AI on the 15th, OpenCode on the 16th and 17th, Railway on the 20th, Warp on the 23rd, CommandCode on the 28th, and Northflank on the 30th. The Cloud Security Alliance connected the incidents as well. Wiz is careful to say there is no single vulnerability or dependency linking them, and that the attackers have not been publicly tied to one actor.

What the reports share is the target. Not source code. Not customer models. Not training data. Environment variables — specifically the ones that hold keys to paid AI services.

The economics of a stolen key

This is where the story stops being a security roundup and becomes a business model. A stolen model key is not worth much to a thief who has to use it personally. It is worth a lot to a thief who can resell access at a discount, because the buyer gets frontier models at a price below list and the seller never touches a GPU.

Wiz tracked two enterprise Amazon Bedrock keys that it could attribute to the vendor breaches. The first showed unauthorized activity within thirty minutes of the reported credential access: a single call from a residential IP, part of a one-minute sweep testing stolen keys across multiple organizations. It asked for Claude Fable 5.1, failed, fell back to Opus 5.5, and worked.

Then came patience. For several days, each key saw only 14 to 120 calls a day, roughly a minute apart, from the same small Go client. That is a liveness check. Then the flood: Case A made about 800,000 Bedrock calls over thirty hours, roughly 90% successful, with Opus models making up about three quarters. Case B ran about 1.7 million calls over thirteen hours, over a million of them successful. At list price, Wiz estimates the two together ran up more than $100,000.

A ring of metal keys resting on the raised floor of a data center aisle

Wiz also found two operations advertising discounted access to frontier models, and both had tested the stolen keys. The researchers are explicit that they cannot definitively tie those operations to the October breaches. They are also explicit that the use of the same keys by multiple actors suggests a functioning market.

How the doors opened

The entry points were different every time, which is the part that should worry anyone building a platform.

Fireworks AI reported unauthorized use of internal credentials, with a third party reaching environment variables used in evaluation jobs in a limited number of customer accounts; it believes API keys and tokens stored as secrets were captured. Zeabur, as reported by BleepingComputer, said an attacker used leaked internal credentials on October 14 and 15 and eventually read the platform database — more than 3.2 million environment-variable records and information on 321,810 accounts. It also said its admin tokens had no expiry and had not been rotated since January 2024. Railway's preliminary view was privilege escalation in its build infrastructure. Warp pointed at a third-party vendor dependency. OpenCode pointed at a legacy system. Lightning AI had been hit before, in an earlier supply-chain campaign.

Two of the incidents were attributed by the vendors to attempted distillation — using another model's outputs to train your own. Fireworks, OpenCode and Warp all reported signs of an "agent swarm" attack. That is a strange sentence to write, but it is where we are: automated agents probing automated platforms, looking for the one environment variable that unlocks somebody's inference budget.

Zeabur's outcome is the one to sit with. The company said it is shutting down. Renewals were disabled from December 1, shared-cluster services paused, and all servers off by December 31, with the breach cited as a contributing factor. A hosting platform does not usually die from one incident. It dies when the incident makes every remaining customer ask whether their secrets are safe, and the answer is a paragraph about tokens that had no expiry.

The instruction people get wrong

Northflank told customers on October 30 to assume every environment variable they kept on the platform had been decrypted, and to rotate every secret. Fireworks made the same point in its own notice, and it is the sentence worth taping to a monitor: deleting a stored secret does not revoke the underlying credential.

Removing an environment variable from a platform removes your copy. It does not touch the key at the issuer. The key is still valid, still billable, and — if it has been resold — still in someone else's client. Rotation has to happen at OpenAI, Anthropic, Bedrock, or whoever issued it, and it has to happen before you go looking at logs, because the logs are how you find out what they did, not how you stop them.

The detection guidance is concrete: Bedrock writes InvokeModel, InvokeModelWithResponseStream, Converse and ConverseStream to CloudTrail as management events, logged by default. Search from October 1 for the addresses Wiz published, knowing a hit is a lead rather than proof. Look for global inference profiles requested by a key that normally stays in one region, quota enumeration from boto3 user agents, and failed calls followed seconds later by a switch to another model.

One architectural note that barely got coverage: CloudTrail records the IAM identity behind a Bedrock API key, not the key itself. If every outside platform shares one identity, you cannot tell which platform leaked. Give each one its own.

What this means if you are the customer

Most teams using these platforms are not running anything exotic. A preview environment, an eval job, a prototype that calls a model from a Railway service with the key in an environment variable because that was the fastest way to ship. The exposure is not a design flaw in the application. It is the default.

Three changes are cheap. Set spending caps per key, so a resale operation cannot run a six-figure bill over a weekend. Issue per-project keys, so revocation is surgical. Alert on abnormal model usage — a jump from a hundred calls a day to a hundred thousand is not subtle, and most billing dashboards will show it hours before anyone reads a security blog.

None of that is a fix for the underlying problem, which is that the industry has spent two years making it trivially easy to put a high-value credential in a low-trust place. The platforms are convenient precisely because they handle secrets for you. That convenience is the attack surface. October was the month it got priced.

Related articles