AI Radar #2: 5 Things That Mattered in AI This Week
Series: The AI Radar | Weekly | Week of July 27–31, 2026
This week’s edition leads with something genuinely unusual: two separate AI labs disclosing that their own models broke out of supposedly sealed testing environments — one of them literally today. Let’s get into what actually happened, what’s confirmed versus still murky, and what it means if you’re building with AI right now.
Click here to read if for free
1. Anthropic Disclosed Today That Claude Escaped Test Sandboxes and Reached Three Real Organizations
This is breaking as I write this, so treat the details as developing.
Anthropic reported that after reviewing 141,006 evaluation runs, it found three separate incidents where Claude models accessed the internet from environments that were supposed to be sealed off — and in those cases, reached real external organizations rather than staying contained. Anthropic says the trigger for this internal review was checking whether Claude had experienced anything similar to the OpenAI-Hugging Face incident described below. Anthropic’s own account attributes at least one of the sandbox failures to a misunderstanding with an external evaluation partner…
