Vibeleaderboard
← All Intel
Intel / article

We got admin access to Baseten's production GitHub in 25 minutes

Source
strix.ai
Author
bearsyankees
Date
From the Daily Brief

Strix's autonomous pentesting agent found a public Harbor container registry belonging to Baseten, pulled a container image from it, and extracted a live GitHub personal access token with admin rights on Baseten's production repositories, all in 25 minutes with no human in the loop. The chain is mundane by security standards: a misconfigured public registry, a credential baked into an image, and a token with far more scope than the image needed. What is new is the speed. An agent that can recon a public registry, decide what to pull, extract a credential, and confirm its scope end to end in under half an hour turns a finding that might once have taken a human researcher a day into something that runs unattended. Baseten is one example; the same registry-to-credential pattern is common wherever container images are built with secrets baked in rather than injected at runtime. Teams that have not audited what their CI publishes to public registries, or that still bake long-lived tokens into images, now have a concrete demonstration of how fast that mistake gets found.

Read the 2026-09-16 Brief →

Context

A publicly downloadable container image can expose access credentials through its build history. In Strix's account, its security found a live credential in an old image belonging to provider Baseten. Read-only checks showed administrative permissions on several GitHub repositories.

The credential dated to a 2023 build and remained usable when discovered in July 2026. Strix says it reported the finding without modifying repositories, and Baseten restricted the registry project and rotated the credential. The failure illustrates why removing a secret from current code is insufficient when downloadable build artifacts still contain a working copy.

Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
Read the source www.strix.ai
Recommended reads
Comments

Checking sign-in…

Loading comments…