Gergely Orosz says that in infrastructure circles it is an open secret: companies are being worked over by what look like Chinese attackers using abliterated models, AI with its safeguards stripped out. What they want is narrow. AI provider keys, most likely to resell at a discount. Break into one platform that holds thousands of customers’ keys and the spend on an offensive model pays for itself.
Railway, a cloud provider, was one of the platforms hit, and its postmortem is the most detailed account yet of how such an attack runs. On 18 October the attacker opened dummy accounts and tried exploits that were blocked. Then came three phases. Five steps in 11 hours to get Railway’s monorepo: a file-read bug in the build system, stolen build credentials, a Temporal cluster holding build histories, the code itself. Four steps in 10 hours to an SSH key in the cloud environment, through a misconfigured OpenID Connect binding in the deploy pipeline. Three steps in three hours to customers’ AI keys, decrypted by the company’s own key management service.
Railway’s team was alerted eight minutes after the second phase began, and revoked the stolen credentials one minute after the attacker used them.
That minute is the story. Jake Cooper, Railway’s chief executive, told Orosz the margin was down to between one and eight minutes. Six hours after downloading the code, the attacker had a plan for reaching the keys. Evicted, it came back the same night to probe for new holes.
Here is the rope. The models Railway could use for defence, from OpenAI and Anthropic, refuse cybersecurity work, defensive work included. Railway had no access during the incident to Project Glasswing, which would have lifted some of those limits. It has access now. Orosz notes that many infrastructure companies do not. The attacker, running a model with no limits at all, never stood in that queue.
Two of Railway’s choices are worth stealing. It kept hashes of customer secrets, so it could send the hashes to the AI providers and have the keys revoked without anyone handling the keys themselves. And rather than patch the build system that let the attacker in, it halted builds and moved every customer to a new one it had nearly finished, with each build in its own throwaway virtual machine. The lessons Orosz draws are blunt: assume the attacker already has your source code, shut down old systems, and answer odd activity in minutes, not hours.
The coda is the other rope, the unwritten one about not selling against a rival’s outage. On the day of the hack Northflank published a post on how to migrate off Railway. Nine days later, Orosz writes, Northflank was hit in a near-identical way. By its own summary the first decryption of customer secrets came between 02:39 and 03:08 UTC on 30 October and the first alert at 11:43 UTC. Orosz counts nine hours to detect and 11 to notify, and says that unlike Railway it did not tip off the AI providers. Railway’s chief executive, he adds, texted Northflank to offer help.