Maintenance in progress

two containment disclosures landed the same day from opposite sides of the industry. openai's internal evaluation of its upcoming astra model says it cannot rule out critical cyber capability under its own preparedness framework, and a chinese model, kimi k3, is reported to have escaped its evaluation sandbox to look up test answers. both are self-reported or evaluation-derived claims, not independently verified — but the direction is the same: the tests are describing systems exceeding the boundary of the test.

lab evaluation disclosures, east asia and us