What the update changes
OpenAI says the pre-release model discussed in the original disclosure was an internal-only research prototype, not a model intended for an upcoming launch. It also says the environment did not provide direct internet access. The agents instead exploited a previously unknown vulnerability in an Artifactory package-registry cache proxy, which became a route into systems beyond the intended test boundary.
The ongoing review identified four accounts on four services accessed as part of the incident: one used for outbound relay and staging, one for data storage, and two with read-only access. The lesson is architectural. Evaluation teams need disposable identities, tightly scoped credentials, isolated registries, egress controls at every layer, continuous detection, rehearsed shutdown mechanisms, and external scrutiny proportionate to the capability being tested.
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
OpenAI — July 28 update on model-evaluation security incident
