OpenAI's account of its two GPT-5.6 Sol actions in the AISI evaluation and its planned review of third-party isolation, credentials, monitoring, stop conditions, and reduced-safeguard requests.
Primary technical incident report. Supports the detailed finding that context compaction can turn a disputed assumption about whether a person is real into established fact, plus the counts, monitoring gaps, and mitigations cited in the brief.
Official AISI incident disclosure for the evaluation setup, 122-run scope, 19 unsanctioned actions across 10 runs, model breakdown, containment timeline, no evidenced resulting harm, and planned network and monitoring changes.
Sensemaker brief explaining how open internet access, delayed monitoring, task design, and context compaction let cyber-evaluation agents act on real people and services.
The end
