Railway breach exposes who pays for stolen AI credentials and US$740,000 in usage

A breach involving cloud hosting platform Railway has exposed the financial risks associated with stolen AI API credentials, with attackers reportedly generating approximately US$740,000 in AI usage charges using keys taken from customers.

The incident highlights an emerging problem as businesses increasingly rely on third-party platforms to store API credentials and connect applications to AI services. When attackers obtain those credentials, they can use them to access AI models, generate substantial usage charges, and potentially distribute stolen access through intermediaries.

Railway said the attackers specifically targeted credentials belonging to AI providers, potentially using the stolen keys to access models for training and model distillation. The investigation found that attackers decrypted stored environment variables belonging to no more than 1.85% of Railway workspaces, affecting approximately 100,000 workspaces.

Database logs showed queries targeting variables matching AI provider API keys, with attackers prioritising accounts containing larger numbers of stored variables.

Stolen AI credentials fuel unauthorised usage

Environment variables are commonly used by developers to store application configuration settings, including database credentials, API keys, and other secrets required for applications to operate.

AI provider API keys function similarly to passwords, granting whoever possesses them access to model services associated with the key owner’s billing account. If attackers obtain these credentials, they can generate requests and leave the legitimate account holder responsible for the resulting usage charges.

Railway believes the stolen credentials may have been distributed through intermediary services referred to as transfer stations. These services can relay or resell access to AI models, potentially allowing stolen keys to reach users who are disconnected from the original attack.

The suspected activity also involved model distillation, a process in which large volumes of model responses are collected to help train another model. This could provide attackers with a way to exploit stolen credentials for AI training workloads rather than simply using them for individual applications.

AI providers take different approaches to refunds

The financial consequences of the breach have varied depending on which AI provider issued the compromised credentials.

AWS Bedrock and Google Gemini have reportedly begun issuing refunds under their existing fraud policies. OpenAI and SpaceXSI are reviewing claims individually through a shared channel established with Railway.

Railway said it would cover costs associated with Anthropic and OpenRouter where providers would not reimburse customers, while discussions with Anthropic continue. OpenRouter has maintained its no-refund policy outlined in its terms of service, leaving Railway to absorb those costs.

The approach represents a change from Railway’s initial response on October 26, when the company told customers it could not reimburse unauthorised spending associated with compromised API keys. Its subsequent incident report outlined a commitment to cover losses where providers declined refunds.

The differing responses highlight how financial responsibility for stolen AI credentials remains inconsistent across providers, particularly when unauthorised usage involves third-party platforms and cloud infrastructure.

Railway resets compromised credentials

The breach also exposed linked-account tokens belonging to 1,852 users, including credentials associated with AI agents and GitHub OAuth integrations.

Railway said it reset the affected tokens, notified users, and found no evidence that stolen credentials had been used to access Railway accounts through unauthorised logins.

The company has filed a complaint with the FBI’s Internet Crime Complaint Center and is working with an external forensics firm to investigate the incident.

AI credential theft targets multiple platforms

Railway described the incident as part of a broader series of attacks targeting platforms that store AI credentials. Similar disclosures involving OpenCode, Warp, and Fireworks AI indicate that attackers are increasingly targeting infrastructure that provides access to AI models.

Fireworks AI disclosed two related incidents involving the exposure of internal credentials and customer information.

The first incident, identified in October, involved misused internal credentials that exposed secrets associated with evaluation jobs for a limited number of accounts. A subsequent breach involving a third-party cloud provider allowed attackers to access Fireworks’ development machines, exposing customer account and billing information alongside certain secrets, datasets, and incomplete model weights.

Fireworks said evidence suggested a financially motivated group was selling access to large language models through illicit routing gateways, a practice known as LLMjacking. This involves hijacking access to AI services and monetising stolen credentials or usage capacity.

The attacks demonstrate how AI infrastructure can become a target in its own right, particularly when credentials provide access to valuable models and potentially expensive compute resources.

Notion affected through third-party AI provider

The Fireworks incident also affected Notion, which had used the platform to classify conversations for its Notion AI Insights feature.

Notion said fewer than 1% of its workspaces may have been affected, with exposed information including AI chat transcripts. For most affected workspaces, the transcripts were not labelled with the identities of individual users or their workspaces.

Notion disconnected its systems from Fireworks and deleted its data from the provider by November 3. It subsequently disabled Insights, stopped serving open-weight models through Fireworks, and advised affected customers to rotate passwords or API keys they had entered into Notion AI.

The incident did not involve attackers directly breaching the affected customers’ own systems. Instead, the exposure occurred through a third-party provider operating between customers and the AI services they used.

Fireworks has engaged Palo Alto Networks’ Unit 42 to investigate and temporarily suspended certain dataset upload and download capabilities while its investigation continues.

Third-party AI infrastructure creates new security risks

The Railway and Fireworks incidents demonstrate how organisations can face security and financial exposure through external platforms that manage AI credentials, workloads, and data.

As companies expand their use of AI agents, API integrations, and third-party model providers, credentials are increasingly distributed across development environments, cloud platforms, and automated workflows. A compromise of any intermediary could expose access to AI models, customer information, and other sensitive resources.

Beyond data exposure, stolen AI credentials can create direct financial losses through unauthorised model usage, particularly when attackers use automated workloads capable of generating substantial token consumption.

Securing AI infrastructure therefore requires more than protecting individual applications. Organisations must also manage credential storage, limit API permissions, monitor usage anomalies, rotate exposed keys, and assess the security of vendors handling AI workloads.

The incidents highlight how the growing AI ecosystem is creating new attack surfaces, with credential theft increasingly capable of generating both data breaches and significant financial consequences.

Source: https://www.artificialintelligence-news.com/news/stolen-ai-keys-railway-breach-who-pays/

Facebook
Twitter
LinkedIn

Related Posts

Leave a Reply

Your email address will not be published. Required fields are marked *