Wiz links AI platform breaches to stolen access tokens

Hackers are increasingly targeting AI platforms to harvest and monetise AI access tokens, with several high-profile breaches disclosed over the past few weeks. SecurityWeek reports that incidents affecting Chainstack, WorkOS, Northflank, Trigger[.]dev and others show intruders seeking stored customer secrets and credentials. Wiz’s analysis attributes the surge to the rise of LLM-jacking and AI broker infrastructure, which makes stolen tokens readily monetisable.

The attacks appear to be driven by access to credentials stored across intermediary hosting platforms, evaluation services and developer consoles, enabling attackers to access inference accounts for multiple downstream customers.

In one strand of activity, Artex — a penetration-testing framework frequently used by threat actors — is cited as enabling automated attack workflows. Chainstack disclosed that on 26 October attackers gained entry to a development environment and parts of production, exposing a database with customer account information such as organisation names, member emails, and password hashes; node credentials may also have been exposed.

WorkOS reported that on 30 October an internal sales-facing app was compromised, exposing business contact details, customer communications, sales data and customer-use information. The report notes the broader pattern of intrusions across numerous AI providers, with intruders leveraging compromised credentials or third-party dependencies and sometimes exploiting specific vulnerabilities, while some claims reference Chinese-APT activity linked to Artex-enabled campaigns. The Cloud Security Alliance emphasises that compromises at this intermediary layer can expose keys for many inference accounts at once.