Black October is the name Wiz gave to a run of at least nine breaches at AI developer platforms during October 2026: Fireworks AI, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode and, on October 30, Northflank. In almost every case the attackers went after stored environment variables, above all the API keys for OpenAI, Anthropic, Google and Amazon Bedrock, and Wiz traced stolen keys to resale sites that sell cheap model access. If you kept an AI key in any of these platforms, revoke it with the provider that issued it, create a new one, redeploy, and then check that provider's usage logs from early October onward. Deleting the secret from the platform does not stop anyone who already copied it.

This guide lists what each platform has said, explains how stolen keys get used (the practice is called LLMjacking), and walks through the checks and settings that make a stolen key worth less next time. It is built from Wiz's Black October report, the vendors' own notices (Northflank status page, Railway report, Zeabur report, Fireworks bulletin) and provider documentation, first read on October 31 and rechecked on November 14. Several investigations are still open.

What's new (November 14, 2026): Railway's full incident report (November 10) says an unknown flaw in its build infrastructure let the attacker partly escape filesystem isolation, then decrypt variables for at most 1.85% of workspaces (about 100,000). Its compensation bulletin puts customer losses at about $740K or more in AI tokens and says Railway will pay unauthorized AI usage that providers do not refund; it asks for claims by November 18. Cursor's November 13 update says its own breach began with a Cursor access key that Fireworks exposed. Northflank (November 6 update) has found no evidence that customer workloads were accessed and still has no root cause.

Which AI platforms were breached in Black October

Wiz uses "Black October" for this cluster of AI developer platforms (separate from the LiteLLM 1.107.0 release). Dates are disclosure dates.

PlatformDateWhat the vendor says was exposed
Fireworks AIOct 6Environment variables used in evaluation jobs, including API keys and tokens, in a limited number of accounts. Affected customers were emailed.
AndromedaOct 9Wiz says Andromeda called it part of an "industry wide attack"; no detail published.
ZeaburOct 14 to 15The platform database: 3,218,965 environment variable records and 321,810 accounts, plus credentials inside 865 database containers. Zeabur is shutting down on December 31, 2026.
Lightning AIOct 15Wiz says Lightning attributes the root cause to credential access.
OpenCodeOct 16 to 17Third party provider keys saved in OpenCode Console by 1.4% of Console users, per BleepingComputer; some reported unauthorized charges of up to $1,700.
RailwayOct 20 to 21Variables for at most 1.85% of workspaces (about 100,000), decrypted after a build infrastructure flaw. It asks everyone who signed up before October 21 to rotate, and it covers unrefunded AI spend (claims by November 18).
WarpOct 23Agent trace and telemetry data from an attack on a vendor, affecting roughly 1% of users by the CEO's estimate, per BleepingComputer.
CommandCodeOct 28Wiz says CommandCode attributes it to an attempt at distillation.
NorthflankOct 30Secrets held in its control plane; users report AI keys being used.
Trigger.devNov 1Environment variables of a subset of organizations; impacted owners were emailed. Not on Wiz's list, which came out before it.

No single vulnerability ties these together, per Wiz; the shared target is environment variables that grant AI access. Railway wrote that the attack on it "was attempting to obtain AI provider API keys to perform distillation attacks," and Zeabur saw its attacker search everything it read for directly usable keys, "especially keys for AI services." Fireworks, OpenCode and Warp all described signs of an "agent swarm," meaning attackers driving many AI agents at once.

Northflank breach: what happened and what to do

Northflank runs containers, databases, jobs and builds. Its status page timeline (Eastern time):

  • 7:43 AM ET: a Wiz threat alert fires; engineers review it a minute later and ship network policy patches at 8:02, 8:50 and 9:05 AM.
  • 9:40 AM ET: "Rotate environment variables immediately."
  • 9:58 AM ET: "Assume all environment variables have been decrypted. Disable/rotate all keys."
  • 11:12 AM ET: secret rotation for Postgres and MongoDB addons is enabled for all users, and "multiple users are confirming that their AI keys have been used."
  • 7:38 PM ET: users report unauthorized use of AI API keys; Northflank says this may fit the pattern in Wiz's report.
  • 9:04 PM ET: audit logs are switched on for every customer, free accounts included.
  • October 31: secret rotation arrives for MySQL, Redis, Minio and RabbitMQ addons.
  • November 2, 6:09 AM ET: the attack vector was contained within an hour of the first alert, independent forensic auditors are reviewing it, running workloads were unaffected, and customers on dedicated or forward deployed control planes are unaffected. A root cause analysis will follow.

The Hacker News thread passed 445 points within a day. Northflank now asks you to rotate both integration secrets and project secrets, starting with AI provider keys, then cloud credentials, payment keys, private keys and database passwords (use the addon rotation feature for those). Where it could, Northflank revoked access tokens, sessions, Git integrations and cross account IAM roles, so sign in again and reconnect them. Then read the audit log for actions you did not take.

What attackers do with a stolen AI key

Wiz followed two enterprise Amazon Bedrock keys that had been used from a breached vendor's infrastructure. The pattern is worth knowing because it tells you what to look for:

  1. A test within 30 minutes. Each key got a single InvokeModel call from a residential IP address, part of a one minute sweep testing stolen keys across many organizations.
  2. A quiet period. For days, 14 to 120 calls a day from the same small Go client, which Wiz reads as a job checking that the keys still worked.
  3. Mapping. A boto3 script listed Bedrock and EC2 quotas and probed which models each key could reach.
  4. The flood. One key made about 800,000 Bedrock calls in 30 hours, about 90% successful, before it was fixed. The other made about 1.7 million calls in 13 hours, over a million successful. Wiz estimates the two cases alone likely ran past $100,000 at list price.

Wiz linked the keys to two sites reselling discounted model access. That is LLMjacking: someone else runs up your bill and sells the access.

The useful detail is why the second key did less damage per hour. Its IAM policy allowed Bedrock only through us.* inference profiles. The reseller pool asked for global.* and failed for four hours, and once it switched, the account's regional quotas throttled most of the rest. The first key allowed global.*, so its load spread across regions and only about 6% of calls were throttled. Scope and quotas did not stop the theft, but they capped the bill.

How to check whether your AI keys were used

Start with the window. Zeabur advises reviewing bills from October 14 onward, and further back for some of its servers, because the first known use is not proof of the first leak. For keys stored on more than one platform, go back to October 1.

  • OpenAI: the Usage page shows spend and requests; compare it with what your team actually ran. OpenAI's key safety guide says to rotate a leaked key from the API Keys page and contact support about misuse.
  • Anthropic: usage and cost are reported per workspace, in the Console and through the Usage and Cost Analytics API, so a key in its own workspace is easy to audit.
  • Amazon Bedrock: AWS documents that InvokeModel, InvokeModelWithResponseStream, Converse and ConverseStream are logged to CloudTrail as management events by default, with the caller's IP address and user agent. Recent events show in Event history even without a trail.

For a quick look at Event history from the AWS CLI (we did not run this one against a live account):

aws cloudtrail lookup-events --lookup-attributes AttributeKey=EventName,AttributeValue=InvokeModel --max-results 50

If you have a trail delivering to S3, this jq filter summarizes Bedrock calls in the unzipped log files by source IP, access key and client. It skips the extra stream completion events AWS records after streaming calls, which AWS warns would double your count. We tested it on sample records.

jq -r '.Records[] | select(.eventSource=="bedrock.amazonaws.com" and .eventType=="AwsApiCall")
  | [.sourceIPAddress, .userIdentity.accessKeyId, .userAgent] | @tsv' *.json | sort | uniq -c | sort -rn

Anything from an address you do not own is a reason to revoke now. Zeabur lists the user agent Boto3/1.43.111 as one its attacker used, and Wiz publishes four IP addresses seen testing stolen keys.

Rotate in the right order

Fireworks puts it plainly: "Removing a stored secret alone does not invalidate the underlying credential." The order that works:

  1. Revoke at the issuer first. Disable the old key at OpenAI, Anthropic, Google, AWS or whoever issued it.
  2. Create the replacement with tighter scope (next section).
  3. Update the platform and redeploy, since running services keep the old value until they restart. Railway's status page tells customers to rotate, revoke the old credentials and redeploy dependent services.
  4. Rotate what the platform does not show you. Changing a database password in an environment variable does not change it inside the database. Zeabur also found that 602 of the 607 n8n instances it checked kept their encryption key in an environment variable, so every credential saved inside n8n leaked too: replace the n8n encryption key and recreate those credentials.

If you are not sure which secrets you put where, this small script lists the names, never the values, of secret looking variables in every .env file under a folder. Run it in your repositories and compare the list with what is stored on each platform.

#!/bin/sh
find "${1:-.}" -name '.env*' -type f 2>/dev/null | while read -r f; do
  sed -n 's/^[[:space:]]*\(export[[:space:]]\{1,\}\)\{0,1\}\([A-Za-z_][A-Za-z0-9_]*\)=.*/\2/p' "$f" |
    grep -E 'KEY|TOKEN|SECRET|PASS|OPENAI|ANTHROPIC|GEMINI|BEDROCK|AWS' | sed "s|^|$f: |"
done

Connection strings such as DATABASE_URL carry passwords too; add URL to the pattern if you use them.

Make the next stolen key worth less

You cannot stop a platform you rely on from being breached, but you can decide what a stolen key is able to do.

  • Give keys an expiry. OpenAI project keys can be created with an expiration date, and admins can enforce a maximum key lifetime for the organization or a project.
  • Set hard spend limits, not only alerts. OpenAI says "spend alerts alone don't stop API traffic" and recommends monthly limits with hard enforcement. On Anthropic, workspaces take their own monthly spend limit and per model rate limits; the Default Workspace cannot have a spend limit, so put deployed keys in a named workspace.
  • Allowlist IPs where you can. OpenAI's IP allowlisting rejects requests from other addresses even with a valid key.
  • Restrict Bedrock to the regions and profiles you use and keep quotas close to real usage. That is what limited Wiz's second case.
  • Skip long lived keys where possible. OpenAI supports workload identity federation, which swaps a trusted identity for a short lived token; on AWS, roles do the same job.
  • Keep keys in one place you control. A self hosted gateway such as LiteLLM lets apps on outside platforms hold a gateway key instead of raw provider keys. Our LLM gateway vs OpenRouter guide covers the tradeoff, including that the gateway itself becomes a target. Zeabur noted that projects which had already moved their variables off its platform were outside the scope of its database read.
  • Run agents in sandboxes with short lived credentials. Our E2B vs Daytona vs Modal comparison covers the options.

LiteLLM and the tampered Claude Code extension in Cursor's marketplace taught the same lesson this month: any tool or platform that holds your keys is part of your attack surface.

FAQ

I did not get an email. Am I safe?

Not necessarily. Fireworks, OpenCode and Trigger.dev say they emailed the customers they identified, but Railway asks everyone who signed up before October 21 to rotate, and Northflank and Zeabur tell all users to rotate. Rotating a key takes minutes, so do it for any key stored on these platforms.

What is LLMjacking?

LLMjacking is the use of stolen cloud or AI API credentials to run model calls on someone else's bill. Wiz found the stolen Black October keys tested within 30 minutes and later used for hundreds of thousands of calls, with links to sites reselling discounted model access.

Is Northflank safe to use now?

As of November 2, Northflank says the attack was contained within an hour, it has patched and hardened the platform, and outside auditors are reviewing it, but the incident is still open and there is no root cause analysis yet. Rotate everything first, then decide once that report is out.

Does deleting a secret from the platform revoke it?

No. The key keeps working until the provider that issued it revokes it. Revoke at the source, create a new key, then update the platform and redeploy.

Was the OpenCode CLI affected?

The reported exposure involved provider keys saved in OpenCode Console, the hosted service, for 1.4% of Console users. If you only run the OpenCode CLI with keys kept on your own machine, those keys were not stored in Console. Check your provider usage anyway if you ever linked a provider there.