AnalysisNov 10, 20266 min read11 sources

Black October's AI-key thefts are confirmed, but no platform has confirmed Wiz's Bedrock abuse link

In one minute

Six platforms' own pages confirm stolen secrets and env vars in October. Wiz's two flooded Bedrock keys and reseller trail are plausible but rest on Wiz alone, and no vendor has named one actor.

On October 30 Wiz published 'Black October', grouping nine AI developer platform incidents disclosed in October (Fireworks, Andromeda, Zeabur, Lightning AI, OpenCode, Railway, Warp, CommandCode, Northflank) and arguing that agent-assisted intrusions hunting stored environment variables are feeding an LLMjacking resale market. Its new evidence is two enterprise Bedrock keys it attributes to one vendor breach that were validated within 30 minutes and then used for millions of calls.

For builders this matters because PaaS, eval and coding-agent platforms hold provider keys on behalf of many customers, so one platform breach becomes many inference bills. The open question is how much of Wiz's cross-platform and downstream attribution the affected platforms themselves confirm.

What we checked

  • Which affected platforms independently confirm the initial access paths, exposed credential scope, and linkage to the downstream Bedrock-key abuse described by Wiz? partly confirmed. Zeabur, Fireworks and Railway confirm initial access paths and Fireworks, Railway, Lightning and Zeabur confirm exposed environment variables or credentials; Andromeda has no root cause. No platform confirms the Bedrock linkage: Wiz does not name the source vendor, and only Northflank reports user-observed AI key misuse while citing Wiz for similar IPs. Zeabur's attacker IP 23.106.46.239 matches an address Wiz saw checking the keys. [2][3][4][6][5][7]
  • Does Wiz tie the Case A and B Bedrock keys to a named vendor breach or to the resellers it identifies? refuted. Wiz says the keys trace to 'a compromised vendor' without naming it, and states it cannot establish the relationship between the resellers and the initial compromise or the flood. [1]

What we found

  • Wiz's core new evidence is downstream: two enterprise AWS Bedrock keys that it says were previously used heavily from a compromised vendor's infrastructure were first tested within 30 minutes of that vendor's reported credential access, then flooded. Case A saw roughly 800k Bedrock calls over 30 hours with about 90% success; Case B saw about 1.7 million calls over 13 hours. Wiz puts the two cases at over $100k at list price. [1]
  • Wiz itself hedges the actor link. It does not name which of the nine vendor breaches the two keys came from, says it cannot establish the relationship between the reseller operations it found and the initial compromise or the flood, and states that its IOC list is not directly tied to the Black October attacks. [1]
  • On initial access, the platforms' own pages confirm credential-based or infrastructure entry but differ in detail. Zeabur gives a full account: leaked internal credentials, stored in plaintext, opened its shared cluster and database. Fireworks reports unauthorized use of internal credentials, Railway a privilege escalation in its build infrastructure, while Andromeda says it has not established a root cause. [2][3][4][5]
  • On credential scope, the confirmations are consistent: what was exposed was customer-stored secrets and environment variables. Fireworks says environment variables including API keys were captured in some accounts; Railway bounds potential exposure to at most 1.85% of workspaces (about 100,000); Lightning says customer credentials were used to reach environments and connected third-party services. [3][4][6]
  • The strongest independent corroboration of a link is infrastructure overlap and Northflank's status page. Zeabur's report names 23.106.46.239 as an attacker address from October 12, and Wiz says the same address later checked the Case A and B keys. Northflank reports users seeing unauthorised use of AI keys supplied to its platform and calls this possibly consistent with the pattern, pointing to Wiz's report for similar IPs. [2][1][7]
  • The campaign framing is contested. CSA, writing on October 23, said no vendor had linked the incidents to one actor and that the campaign framing came from Railway alone; Railway says the attacker sought AI provider keys for distillation. Andromeda calls its incident part of an industry-wide attack. Wiz itself says no single vulnerability or dependency links the attacks. [8][4][5][1]

What it means

Treat any AI provider key stored on a third-party platform that disclosed an October incident as stolen: revoke it at the issuing provider, not just in the platform's secret store, and check provider usage logs for the pattern Wiz describes (a single validation call soon after exposure, low-volume keep-alive checks, then a quota probe and a flood). On Bedrock, scope keys away from global.* inference profiles where you can; Wiz's Case B shows regional restriction and quotas capped the damage.

Do not read Wiz's report as proof of one coordinated actor. The confirmed facts support a narrower conclusion: several platforms lost customer secrets in the same month, and stolen AI keys monetize fast through resale. Budget alerts, per-key quotas and short-lived credentials are the controls that limit cost regardless of who did it.

Competing readings

  • Several separate intrusions, possibly by different actors, went after AI keys because a liquid resale market makes them easy to monetize. On this reading Black October is a pattern, not one campaign (likely, about 60%). What would tell it apart: The platforms report different root causes: leaked internal credentials at Zeabur and Fireworks, privilege escalation in Railway's build infrastructure, and no root cause yet at Andromeda [2][3][4][5]; Wiz cannot tie the reseller operations to the initial compromise, and its IOCs are not directly linked to the Black October attacks [1].
  • One coordinated actor or crew ran the intrusions and fed the stolen keys into resale (unlikely, about 30%). What would tell it apart: Wiz links the abused Bedrock keys to a compromised vendor's infrastructure, which fits a breach-to-abuse pipeline [1]; Several platforms lost secrets in the same month, including environment variables that held API keys [2][3].
  • The downstream Bedrock abuse is unrelated to these platform breaches, and Wiz over-attributes it (very unlikely, about 10%). What would tell it apart: Wiz does not name the source vendor, so no affected platform can be shown to have confirmed the link [1]; Wiz's IOCs are not directly linked to Black October, and the reseller trail is not tied to the compromise [1].

Base rate

The New Runtime archive holds only 3 earlier research stories about companies or products from the last 12 months that are at least 3 months old: too few (minimum 5) for a base rate. The archive has only 3 comparable cases, 2 of them matches. That is too few for a base rate (the minimum is 5), so this analysis does not lean on it.

What would change this view

  • A second affected platform publishes IOCs or forensics naming the same operator or infrastructure as another disclosed incident (signal): The single-actor campaign reading would become more likely than the converging-market reading.
  • The vendor behind Wiz's Case A and B keys disputes that its breach exposed those keys (signal): Wiz's downstream attribution would be weaker, which raises the unrelated-abuse reading.
  • Resale sites stop offering discounted frontier-model access, or providers tighten key validation and pricing for cross-region routing (pricing signal): The resale market that drives the pattern would be weakening.

Still open

  • Which vendor breach did Wiz's Case A and B keys come from, and will that vendor confirm it?
  • Will any platform beyond Zeabur publish IOCs that overlap with the key-checking addresses Wiz lists?
  • Did AWS or model providers detect and revoke keys from these breaches proactively, as Railway's notifications intended?
  • What were the initial access paths at Andromeda, Warp, OpenCode, CommandCode and Northflank?

Evidence / sources

  1. 1wiz.io/blog/black-october-for-ai-developer-platforms
  2. 2zeabur.com/blogs/october-2026-security-incident
  3. 3fireworks.ai/security/bulletins/october-2026-security-incident
  4. 4blog.railway.com/p/preliminary-report-october-security-incident
  5. 5andromeda.ai/security-updates/october-2026-security-incident
  6. 6lightning.ai/blog/oct-2026-incident
  7. 7status.northflank.com
  8. 8labs.cloudsecurityalliance.org/research/csa-research-note-inference-provider-platform-chain-campaign
  9. 9zeabur.com/blogs/zeabur-is-shutting-down
  10. 10bleepingcomputer.com/news/security/warp-discloses-breach-exposing-ai-agent-logs-user-telemetry-data
  11. 11thehackernews.com/2026/10/opencode-console-breach-may-have.html
All Analysis

Next reads

Read next

Open the graph
  1. 01archiveAnalysisOpen the latest authorial analysis.
  2. 02sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract