Railway's Two-Flaw Chain Let Attackers Run Up $740,000 in Stolen AI Credits
Cybersecurity

Railway's Two-Flaw Chain Let Attackers Run Up $740,000 in Stolen AI Credits

A build-infrastructure escape and a misconfigured trust relationship let attackers decrypt customer secrets at app hoster Railway, part of a wave Wiz calls Black October that is hunting AI provider credentials across developer platforms.

PublishedNovember 15, 2026
Read time5 min read
Share

Two flaws, chained inside a 26-hour window

Railway's November 10 incident report lays out a two-stage attack that began October 20. Attackers first exploited a previously unknown vulnerability in Railway's build infrastructure to escape filesystem isolation and obtain an internal credential. With that foothold, they accessed stored build histories and copied Railway's internal source code, which they then studied to find further weaknesses in the platform they had just breached. That is the part worth underlining: the attackers used Railway's own code against Railway, turning a single foothold into a map of what to look for next.

The second flaw they found was an improperly configured trust relationship that handed them credentials capable of decrypting customer environment variables, the stored secrets every Railway-hosted application depends on. Railway says its engineers fully removed the attackers from its systems by October 21, roughly 26 hours after the intrusion began, and disclosed the incident to customers the following day with instructions to rotate secrets.

What was actually exposed

Railway's own characterization is careful: attackers gained access to decrypted environment variables for at most 1.85 percent of workspaces, roughly 100,000, and the company is explicit that this is a maximum exposure estimate rather than a confirmed count of workspaces whose secrets were actually stolen. Separately, linked-account tokens for 1,852 users were exposed, including GitHub OAuth tokens and AI agent keys, all of which Railway has since reset and notified the affected users about.

Railway reports no evidence that customer Git repositories or customer databases were accessed, which narrows the blast radius to credentials and environment configuration rather than application data or source code. For a developer platform, that is a meaningful distinction, though it offers little comfort to any customer whose AI provider keys were among those stolen and abused.

The attackers wanted AI credentials specifically

Railway says the attackers specifically searched for AI provider credentials during their time inside the platform, though some of their queries also targeted cloud, database, and payment service credentials. The stolen AI credentials generated approximately $740,000 in unauthorized usage charges before Railway and affected providers could shut off access. Railway's working theory is that the attackers wanted the credentials either for resale or to use in AI model distillation, extracting a competitor's model behavior by running large volumes of queries through someone else's paid API access.

That motive matters because it describes a new, specific market for stolen credentials. API keys have always had resale value for compute arbitrage, but distillation gives attackers a second, more durable reason to want access to frontier AI APIs: not just cheap compute, but a way to approximate a model's behavior without paying for the underlying research.

Part of a named campaign, not an isolated incident

Railway links this breach to a wave of attacks on AI developer platforms that Wiz has named Black October. Other platforms reportedly hit in the same campaign include Northflank, OpenCode, Warp, and Lightning AI, all developer infrastructure providers whose customers store API keys and AI credentials as a normal part of using the platform. That pattern, multiple infrastructure providers hit in the same window with the same credential-theft objective, points to a coordinated or at least systematically repeated campaign rather than four unrelated opportunistic attacks.

For any enterprise that builds on top of a developer PaaS, this is the risk model to internalize: your AI provider credentials live on someone else's infrastructure, and that infrastructure is now a named target for a specific class of attacker hunting those credentials across every platform that stores them.

Railway's response and what it tells other platforms

Railway says all exploited vulnerabilities have been remediated, it is working with affected AI providers to reimburse customers for fraudulent usage, and it will personally cover any charges a provider declines to refund. That is a stronger customer-facing commitment than most infrastructure vendors make after an incident, and it sets a bar other developer platforms will likely be measured against going forward.

Railway's own description of the incident, a roughly 26-hour contest between its engineers and attackers using AI-powered agents, is also a preview of what incident response increasingly looks like: not a single breach followed by a single fix, but a running fight where both sides are using automated tooling to move faster than the other. Any CTO evaluating a developer platform vendor should be asking now, before an incident, how that vendor segments build infrastructure from customer secrets, and how fast they can actually detect and evict an intruder.

The budget conversation this forces

A $740,000 fraudulent usage bill, even one a vendor has pledged to cover, is a number finance teams will ask about the moment it surfaces in a vendor security review. Enterprises that have scaled AI usage quickly, often provisioning keys across multiple teams and environments without centralized tracking, are exactly the kind of customer whose compromised credentials could generate charges that size before anyone notices the anomaly in a monthly bill.

The defensive answer is not complicated, just neglected: set hard spend caps and anomaly alerts on every AI provider account, treat API keys stored in any third-party build or hosting platform as a monitored asset rather than a set-and-forget credential, and confirm with each developer platform vendor, in writing, what their reimbursement commitment actually covers before you need to invoke it. A $740,000 anomaly took Railway's own engineers roughly a day to contain once they knew to look, which means any enterprise without comparable spend alerting of its own would likely have noticed the charge only after the invoice arrived.

Tagged#news#security#cybersecurity#breach#cisa#ransomware#zero-day#supply-chain#ai-security#cloud-security#credential-theft#developer-platforms#railway#build-infrastructure#wiz-research#llm-credential-theft