News · DiggingBeagle record

Northflank credential exposure coincides with unauthorized AI inference traffic

A deployment-platform security notice and separate Bedrock traffic investigations show how stolen AI keys can finance unauthorized inference without further access to a model host.

Published
Nov 1, 2026

The report

Northflank instructed customers to consider stored environment secrets compromised after a platform security incident, and its status history reports unauthorized use of AI provider credentials. It published staged rotation guidance for supported databases and addons, but its public disclosures do not establish the exact credential-access path for every affected customer.

Separately, Wiz documented two customer credentials that enabled unauthorized Amazon Bedrock traffic. One key generated about 800,000 requests over roughly 30 hours, and a second generated about 1.7 million requests in 13 hours as the actor adjusted inference profiles and quota tests. These are inference request counts from separate credentials, not compromised tenant counts. Overlap in testing infrastructure does not prove that one actor originally stole both keys.

The meaningful boundary is the credential's authority at the issuing AI provider: it may be billable even after the original host is no longer under attacker control. Immediate provider-side revocation, least-privilege model grants and independent usage audits therefore complement platform-secret rotation. Northflank exposure and the distinct Wiz Case A and Case B preserve the separate evidence.

The next-day response was a series of distinct credential actions

Northflank's public status history supplies more precise remediation timing than the initial incident narrative. It reported customer accounts of AI provider key misuse on October 30, enabled new audit-log access for all customers on October 31, and then documented progressive secret-rotation support for Postgres and MongoDB, MySQL and Redis, MinIO, and RabbitMQ add-ons. This sequence identifies capabilities offered to customers, not proof that every tenant had completed rotation.

The operator-facing response has two steps that should not be merged. First, revoke exposed AI provider credentials at the issuing service and review provider-side usage and billing, because editing a secret inside Northflank does not invalidate a copy of the old key. Second, rotate platform and addon secrets and verify which workloads actually received the replacement values. The status timeline does not establish one confirmed initial intrusion route for each affected customer, nor does it prove that Wiz's two illustrated Bedrock keys originated on Northflank.

Research behind this