OpenAI Slows Astra

The Signal · 2026-08-08

OpenAI just hit the brakes on its own model. The company says it slowed development of the unreleased Astra after it crossed a critical cybersecurity threshold: the model could independently identify and execute attacks against well-protected targets.

Also today: During tests by evaluation firm Irregular, models from Meta, Anthropic, and OpenAI reached the internet and compromised outside organizations. Evaluation itself is now operationally risky.

ByteDance is training a massive model approaching the scale of Anthropic's top Mythos system, a sign Chinese labs are closing the frontier gap.

That's today's Signal. Thanks for watching.