All articles

NewsNovember 10, 20265 min read

Railway: stolen AI keys and $740,000 in fraudulent charges

Railway reconstructs the attack: two flaws, environment variables decrypted for up to 100,000 workspaces and customer AI keys used for about $740,000.

Open laptop with a light blue cloud-shaped padlock in the middle of the screen

In the incident report published on 10 November, Railway reconstructs the attack that led it, on 22 October, to ask all users to rotate their secrets. The new information covers three points: how the attackers got in, how many environment variables they may have decrypted, and what they did with customers’ AI provider keys: about $740,000 in unauthorised usage.

The story is reported by BleepingComputer, The Hacker News and CSO Online; the details that follow come from BleepingComputer and CSO Online, which quote Railway’s report. When a fact is reported by a single outlet, we say so.

The attack chain

According to the report, the attack began on 20 October and unfolded in three steps:

  1. Escape from build isolation. The attackers exploited a previously unknown vulnerability in Railway’s build infrastructure to escape filesystem isolation and obtain an internal credential.
  2. Source code and reconnaissance. With that credential they read stored build histories and retrieved a copy of the platform’s internal source code, which they then analysed to find other routes in.
  3. Misconfigured trust relationship. A second vulnerability, an improperly configured trust relationship, gave them further access to the infrastructure and the credentials used to decrypt customers’ environment variables.

Railway describes a fight of roughly 26 hours: the attacker chained exploits, moved laterally across several systems and probed many routes in parallel, while engineers closed them, also in parallel. The company calls it a “contest of agent speed”. Only BleepingComputer reports that, according to Railway citing Wiz, the attackers used AI agents based on “abliterated” models, meaning models modified to remove their safety refusals, and a harness similar to ARTEX.

The attackers were evicted on 21 October and, according to BleepingComputer, all exploited vulnerabilities have been fixed. No CVEs or exploit details have been published: according to CSO Online, the ongoing investigation does not allow it.

What was exposed

  • Decrypted environment variables: at most 1.85% of workspaces, about 100,000. According to BleepingComputer this is the maximum number of potentially affected workspaces, not the number whose secrets are confirmed stolen.
  • Linked-account tokens of 1,852 users (CSO Online says “more than 1,850”), including GitHub OAuth tokens and, according to BleepingComputer, AI agent keys. Railway reset them and notified the users concerned.
  • According to BleepingComputer, there is no evidence that the attackers downloaded customers’ Git repositories or accessed their databases.
  • According to analyst Frank Dickson, quoted by CSO Online, Railway admits in the report that it cannot be sure it found every account the attacker created.

AI keys, the real target

The attacker’s queries looked first of all for AI provider keys, by name; some also targeted cloud, database and payment service credentials. The stolen keys generated about $740,000 in unauthorised usage. According to CSO Online, some customers saw their keys answering coding questions and running training-data tasks they had never launched. Railway is working with providers on refunds and will pay the charges providers decline to refund; CSO Online mentions key revocation coordinated with GitHub, GitLab, Anthropic, OpenAI, OpenRouter, Stripe and others.

Railway believes the keys were meant for resale or for model distillation, that is, training one model on another model’s answers. As Dickson explains to CSO Online, this takes millions of questions, and a single account asking them stands out: spread across thousands of real customers’ keys, they go unnoticed. According to BleepingComputer, Railway links the attack to the wave Wiz calls “Black October”, which over the past month has also hit Northflank, OpenCode, Warp and Lightning AI.

A useful data point for response times comes from a Wiz investigation quoted by CSO Online: stolen keys were tested within 30 minutes of the theft and flooded with traffic four to seven days later.

What Railway changed

According to CSO Online, Railway shut down potentially compromised internal tooling, rotated all keys and then rotated them again after confirming the attackers were out, moved to a new generation of infrastructure and added mass revocation of Railway app permissions in case of suspicious activity. It is also moving to short-lived credentials and logging every decryption.

ATT&CK techniques

Indicative mapping, derived from the reported facts: T1611 (Escape to Host) for the escape from build isolation; T1552 (Unsecured Credentials) for the internal credential and the decryption credentials; T1496.004 (Resource Hijacking: Cloud Service Hijacking) for the use of AI keys at customers’ expense; T1078.004 (Valid Accounts: Cloud Accounts) for the use of the other stolen credentials. No indicators of compromise have been published.

Impact for organisations

Compared with the October notice, two things change: the damage is no longer theoretical, because the keys were used and someone paid, and timelines are tight, because only minutes pass between theft and first use. For European and Italian companies subject to NIS2, the report is the document to attach to the supplier assessment. For those not using Railway, Dickson suggests asking your platform provider two questions: how are secrets stored, and who can decrypt them?

What to do now

  1. Rotate everything that existed before 21 October: Railway asks users to revoke and reapply sensitive variables stored up to that date. Start with AI provider keys, then cloud, database and payment credentials.
  2. Check AI provider bills and usage logs from 20 October onwards for traffic your teams did not generate, and report unauthorised charges to Railway.
  3. Limit the damage any key can do: hard spending caps, usage alerts and permissions restricted to the models you actually need. In the case Wiz examined, a key restricted to a few models throttled most of the thief’s traffic.
  4. Remove public network access to databases hosted on Railway, as the company advises.
  5. Check the GitHub audit log and authorised apps, and the audit logs of registries such as Docker Hub.
  6. Set alerts for calls from unrecognised addresses to your services and APIs.

How ARXDOME helps

If the secrets included credentials for servers or databases you manage, ARX Sentry collects security logs and events and groups them into families, such as explicit credential use, in the SOC pages. ARX Horizon enriches every connection with country, operator and threat lists, so an access with valid credentials from an unusual address stands out immediately.

Last updated:

Any questions?

Let’s talk about your company’s needs, with no obligation.

Contact us

Related articles