Good morning, and welcome back to Techzip AI Newsletter.

Today’s issue looks at what happens when AI moves from answering prompts to operating inside real systems. Anthropic is using Claude for formal verification across open-source libraries, while Claude Duet tests proactive background work with permission gates. OpenAI says its Patch the Planet initiative helped uncover a serious PHP vulnerability, and separate OpenAI reports show why independent monitoring matters when agents can disable or misreport their own oversight.

We also look at the capital behind frontier AI, Canva Craft, Microsoft MAI-DxO, MiniMax M3.1 Flash, and OO-Harness—the idea that better AI agents may need better orchestration, not only stronger models.