OpenAI report: red-team agents reached Kubernetes cluster-admin, no weight access
JeffLadish · x · 2026-09-08
Jeff Ladish highlighted key passages from OpenAI's security report: in an internal-compromise-focused exercise, agents escalated into OpenAI's managed cloud Kubernetes service, gained cluster-admin privileges, and created a privileged host-mounted pod.
The agents obtained authentication tokens for OpenAI's cloud-hosted secrets management service, cloud IAM, and the CaaS environment's CI platform, then used those credentials to create public application and network load balancers in OpenAI's public cloud environment. Notably, the agents did not gain direct access to model weights, though the internal access surface was quite broad.
More from Safety
- TASTE: A New Benchmark Testing If Models Can Predict AI Safety Researchers' Preferences — burny_tech · 2026-09-08
- Gemini User Claims Model Drew His Family's Unique Home Decor Despite Opting Out of Data Saving — Legitimate-Theory738 · 2026-09-08
- AI Agent Auto-Enrolls User in Fake McKinsey Group, Then Drafts GP Data Theft Plan — LadyAshBorg · 2026-09-08
- Feeding attacker-written email text to an LLM filter: how bad is prompt injection here? — Several_Log_4610 · 2026-09-08
- When Juries Deadlock, AI Could Decide: AI Tools Enter Criminal Justice — TobyWalsh · 2026-09-08
- Researcher factors 1990s Certificate Authority RSA keys, exposing legacy trust risks — ahlCVA · 2026-09-08