Skip to main content
The state of ai impact assessment
IncidentPeople used AI as a toolSeverity S: serious, 4/5Confirmed: a company, official or court statement

Railway breach by AI-assisted attackers exposed customer AI API keys, leading to $740,000 in fraudulent AI charges

Cloud hosting company Railway said unidentified attackers used a zero-day in its build infrastructure to partially escape filesystem isolation. From there they took internal credentials and accessed other internal systems, including credentials needed to access a production database of encrypted customer environment variables, along with information on how to decrypt them. The attackers were mostly after customers' AI API keys. The stolen keys were used for roughly $740,000 in unauthorized AI usage. Railway believes whoever obtained the keys likely sold them, possibly to support model distillation. Railway said it spent 26 hours fighting the attackers, who, according to Wiz, used abliterated models and a harness similar to ARTEX. It estimates that environment variables of up to 100,000 customer workspaces could have been decrypted, but says this is an upper limit, not a confirmed number of compromised accounts. Railway has closed the vulnerabilities, rebuilt affected infrastructure and rotated credentials. It is working with AI providers to refund fraudulent charges and has promised to cover losses the providers refuse to reimburse.

What the AI did

Unidentified attackers reportedly used AI tools during the 26-hour attack: so-called "abliterated" models (AI models altered to remove their built-in safety limits) and an automation setup similar to one called ARTEX. The AI API keys they stole from Railway customers were then used for roughly $740,000 of unauthorized AI usage, which Railway believes may have gone towards copying other AI models.

First reported November 11, 2026 · Added to the register November 16, 2026 · 4 sources

Fraudulent AI charges
$740,000
Users whose linked-account tokens were taken
1852 users
Users with decrypted linked-account tokens
1,852 users
Users whose linked-account tokens were exposed
more than 1,850
Affected
Railway
Country
United States
When it happened
October 20, 2026 to October 21, 2026
First reported
November 11, 2026

What this means for you

Could this affect you?

Yes

If you host on Railway. Railway customers who stored AI provider API keys or cloud credentials as environment variables (settings stored with their apps) on the platform may have been affected. Railway says environment variables of up to 100,000 customer workspaces could have been decrypted, though this is an upper limit rather than a confirmed count.

What to check

  • Rotate any API keys and secrets stored as environment variables on Railway.
  • Review your AI provider bills for usage you did not authorize.
  • Set spending limits and alerts on your AI API keys.
  • Strengthen build isolation and limit internal credentials if you run a hosting platform.

Timeline

  1. October 20, 2026Started
  2. October 21, 2026Ended
  3. November 10, 2026Railway reported the incident to the FBI's Internet Crime Complaint Center. It had initially told affected customers on Oct. 26 that it could not refund charges generated through compromised API keys, a position its Nov. 10 report reversed.[4]
  4. November 11, 2026First reported
  5. November 11, 2026Railway's post-mortem said Amazon Web Services and Google have begun refunds under their fraud policies, OpenAI and SpaceXSI are reviewing claims case by case, and OpenRouter is enforcing the no-refund policy in its terms of service. Anthropic was still in discussions when the report was published.[1]
  6. November 12, 2026Railway is moving to short-lived credentials and logging every decryption.[3]
Every fact and its source (16)
  1. Attack startedOctober 20, about 07:28 UTC
    “the attack started on October 20 at roughly 07:28 UTC”[2]
  2. Attack endedOctober 21, 09:24 UTC
    “ended at 09:24 UTC on October 21, some 26 hours later”[2]
  3. Intruder evictedOctober 21
    “Railway said it evicted the intruder on October 21.”[1]
  4. Intrusion beganOctober 20
    “The intrusion began on October 20”[1]
  5. Attack datesOct. 20–21
    “The attack occurred Oct. 20–21 and involved”[4]
  6. Attack began2026-10-20
    “The attack began on October 20, when hackers exploited a previously unknown vulnerability”[3]
  7. Attackers evicted2026-10-21
    “on October 21, but over the course of their time in the system”[3]
  8. AI tooling used by attackersAbliterated models and a harness similar to ARTEX
    “Railway defending against abliterated models and a harness similar to ARTEX, according to Wiz”[2]
  9. RemediationClosed vulnerabilities, rebuilt infrastructure, rotated credentials
    “Railway said it has since closed the exploited vulnerabilities, rebuilt affected infrastructure, and rotated compromised credentials.”[2]
  10. Refund commitmentWorking with AI providers on refunds; will cover unreimbursed losses
    “It is also working with AI providers to refund fraudulent charges, promising to cover losses that the providers refuse to reimburse.”[2]
  11. Upper limit of workspaces whose environment variables could have been decryptedup to 100,000 customer workspaces
    “environment variables of up to 100,000 customer workspaces could have been decrypted during the incident”[2]
  12. Provider refundsAWS and Google refunding; OpenAI and SpaceXSI reviewing
    “Amazon Web Services and Google have begun refunds under their fraud policies, while OpenAI and SpaceXSI are reviewing claims case by case”[1]
  13. OpenRouter positionNo refunds
    “OpenRouter is enforcing the no-refund policy in its terms of service”[1]
  14. Build histories accessed353 build histories
    “353 build histories: Attackers accessed stored build histories”[4]
  15. Law enforcement reportReported to FBI IC3
    “The company has also reported the incident to the FBI's Internet Crime Complaint Center”[4]
  16. Railway remediationShort-lived credentials and logging every decryption
    “Railway is now moving to short-lived credentials and logging every decryption.”[3]

Sources

  1. App hoster Railway attackers ran up US$740,000 on stolen AI keys
    itnews.com.au · November 11, 2026
  2. Cloud hosting firm Railway says recent breach leads to $740,000 in fraudulent AI charges
    techradar.com · November 11, 2026
  3. Hackers stole AI credentials to get $740,000 in free compute; Railway picked up the bill
    csoonline.com · November 12, 2026
  4. Railway Breach Leads to $740,000 in Fraudulent AI Charges
    esecurityplanet.com · November 12, 2026

How this record is classified. Severity S (4/5): real harm to one organisation, data exposed, or a containment failure. OECD level: incident, an event in which an AI system led to actual harm. Evidence: Confirmed, meaning a company, official or court statement.

The facts in At a glance and in Every fact and its source are quoted from the sources listed. The summary, What the AI did and What this means for you are written from those sources and checked against them automatically before publication. Records reflect what has been disclosed, not everything that has happened. Incidents are found through news feeds and the GDELT Project.

Promotional banner for the Pentest Readiness checklist download