Cybersecurity 6 min read

Railway Says Attacker Decrypted Secrets From Up to 100,000 Workspaces as AI Platform Breaches Reach 16

Railway's full report puts stolen-key spending at about US$740,000. We checked all 16 platforms on Wiz's tracker against their own notices: ten posted one, five only on X.

Priya Nair
Data, AI Governance & Policy Analyst
Published 11 Nov 2026, 1:30 PM (SGT)
Share:
A dimly lit aisle between rows of server racks in a data centre A dimly lit aisle between rows of server racks in a data centre Photo by Akela999 on Pixabay
Advertisement

11 NOV 2026 — Railway, the cloud hosting company, has published its full incident report on the October intrusion. By the company's count, the attacker spent close to 26 hours inside and decrypted stored secrets belonging to at most 1.85% of workspaces, or about 100,000. The report, out on 10 November, says the intruder went looking for AI provider keys and that unauthorised spending on stolen keys was estimated to reach about US$740,000.

Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, which Railway names among the AI providers it asked to revoke stolen keys. Readers should weigh that interest.

Railway is one of 16 developer platforms listed on a running tracker kept by the security firm Wiz, which calls the run of incidents "Black October". We went back to each company's own notice to see what each one actually admits.

Railway's 26 hours

Railway says the attack began at 07:28 UTC on 20 October, when the intruder exploited a previously unknown flaw to partly escape filesystem isolation in Railway's build system. That yielded a build credential, and through it a copy of Railway's internal monorepo. From the code, the attacker found "an erroneously defined OIDC binding configuration" in one of Railway's cloud environments, and at 03:26 UTC on 21 October obtained short-lived cloud credentials through it.

Decryption of customer variables ran from 04:42 UTC until Railway disabled the credential at 07:14 UTC, a window of two hours and 32 minutes. Railway says the attacker was "fully evicted" at 09:24 UTC. From the first exploit to eviction is 25 hours and 56 minutes, by our count from Railway's own timestamps.

Railway does not give its total number of workspaces. If 1.85% is about 100,000, the implied base is roughly 5.4 million (100,000 ÷ 0.0185), though that figure is our arithmetic and Railway calls the 1.85% an upper bound. Build histories for 353 builds were read as well, and the decrypted data included linked-account tokens for 1,852 users, all of which Railway says it reset.

Database logs show the attacker "specifically prioritizing AI provider keys". Railway says it sent revocation requests to eight vendors, among them GitHub, Anthropic, OpenAI, Google and Stripe, and that it will cover token losses itself where a provider declines to refund them. It has also filed a complaint with the FBI's Internet Crime Complaint Center.

Sixteen platforms, in their own words

The table sets out the 16 platforms on Wiz's list as of 10 November. Dates and descriptions come from each company's notice, not from the tracker. A RunPod incident that Wiz says was circulating only on social media is left out.

PlatformDate the company givesWhere it said soWhat it says was exposed
Fireworks AIIdentified 6 Oct; update 9 NovOwn siteEnvironment variables in evaluation jobs; later, billing data and some datasets and incomplete model weights
AndromedaIdentified 13 OctOwn siteAndromeda and customer credentials, limited number of accounts
ZeaburIncident 14-15 OctOwn siteEnvironment variables in the platform database; commands run in 865 database services
Lightning AIIdentified 15 OctOwn siteCustomer credentials used to reach environments and connected services
OpenCode17 OctX post1.4% of OpenCode Console users
Railway21 Oct; full report 10 NovOwn siteVariables for at most 1.85% of workspaces
Warp26 OctCEO's X postA vendor in Warp's infrastructure exploited
Command Code28 OctX postAll API keys revoked as a precaution
Northflank30 OctStatus pageSecrets decrypted; users report misuse of AI API keys
Brev (NVIDIA)Activity 4-20 OctCustomer notice quoted by Tom's HardwareSSH keys, since rotated
Trigger.dev1 Nov; report 6 NovOwn siteEnvironment variables, 0.66% of customers
ChainstackDetected 31 OctOwn siteAccount data including password hashes; node credentials treated as exposed
Kernel4 NovOwn site0.28% of customers; API keys and third-party secrets in logs
BlacksmithDetected 5 NovX postThe system storing CI job logs
Tensorlake8 NovX post by Diptanu ChoudhuryNo impact to customer API keys or sandboxes, the post says
NEAR AIPublished 9 NovOwn siteAccount details of about 100 users, 18 agents, two API keys

Ten of the 16 put a notice on their own website or status page. Five said so only on X, from a company account or an individual's. For Brev we found no public notice from NVIDIA, only a customer email quoted by the press. Four companies give a share of users hit, and all are small: kernel 0.28%, Trigger.dev 0.66%, OpenCode 1.4% and Railway at most 1.85%.

Advertisement

Different doors, no named attacker

No company names who was behind the attacks, and the entry points differ. Zeabur says long-lived admin credentials "were stored in plaintext", and it has announced it will shut down on 31 December, a move it calls an operational decision. NEAR AI blames a remote code execution flaw in SGLang, the open-source inference engine. Chainstack points to a third-party component. Fireworks says the evidence links its case to a financially motivated group selling stolen access to large language models, but names no group.

Wiz says "there is no single vulnerability or dependency linking these attacks". Railway, for its part, says the Wiz research "does not independently attribute the Railway incident", and that it has not established whether the intruder ran the key-resale services where stolen credentials turned up.

The harm also travels downstream. Notion told users that fewer than 1% of its workspaces may have had unattributable Notion AI chat transcripts exposed through the Fireworks incident. For anyone who stored a key on one of these platforms, Fireworks puts the fix plainly: "Removing a stored secret alone does not invalidate the underlying credential." The key has to be revoked with the provider that issued it.

Advertisement
Priya Nair
Data, AI Governance & Policy Analyst

Priya Nair covers AI governance, data protection, privacy, and digital trust topics for RECATOOLS.

View author profile → · Editorial policy

About this byline Priya Nair is a RECATOOLS editorial persona for AI governance, privacy, and digital trust coverage. Articles are produced and reviewed under RECATOOLS editorial supervision.

Corrections policy

Advertisement