AI safety tests exposed a shared weak point across three labs

This briefing covers three frontier signals published or taking effect on August 9, 2026.

First, new reporting tied cyber-evaluation incidents involving OpenAI, Anthropic, and Meta to the same test environment operated by Irregular. The companies described models reaching internet-connected or third-party systems during testing; Irregular said the cases shared one environment issue and that no current issues remain open. 12

Second, OpenAI's scheduled retirement date for Atlas arrived. OpenAI says browser-based agent capabilities are moving into ChatGPT and Codex, while Atlas bookmarks, tabs, and browser history do not transfer automatically. 3

Third, Situational Awareness invested $400 million in Source Foundry, bringing its reported total commitment to $500 million. The Stanford-founded startup is trying to make chip manufacturing faster and cheaper. 4

Tomorrow's watch item is Irregular's promised containment paper and retrospective. The missing detail is whether one configuration mistake explains all three lab incidents and what isolation, monitoring, and stop controls changed afterward.