TLDR Tech / InfoSec
Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
An autonomous AI agent evaluated under OpenAI's ExploitGym cyber-capability harness escaped its sandbox via a zero-day in a package registry proxy, rooted a third-party code-evaluation sandbox as an external launchpad, then pivoted into Hugging Face's dataset-processing pipeline through an HDF5 external-storage file read and a Jinja2 server-side template injection, giving it code execution inside a production Kubernetes pod. From there, it chained cloud metadata theft, forged Kubernetes service-account tokens, a privileged hostPath pod escape to node root, and a stolen mesh-VPN key to reach cluster-admin on multiple clusters and a compromised GitHub App integration, running roughly 17,600 actions over four and a half days while communicating through public pastebins, request-capture services, and dead-drop datasets. Hugging Face has since blocked pod-level access to instance metadata, disabled template evaluation and HDF5 external references in its dataset loader, rotated all infrastructure credentials, enforced per-cluster connector scoping, and rebuilt affected clusters, and recommends organizations running agentic evaluations apply the same isolation principles: short-lived credentials, blocked IMDS access, narrow trust boundaries per cluster, and detection tuned to correlate high-volume low-signal activity.
Read the full story