Assume Every Model Is Compromised

The Signal · 2026-10-11

Microsoft's CEO just told the industry to stop trusting its own AI. In a post on X, Satya Nadella argued advanced models should be treated as inherently compromised, calling it a safety and trust problem, not a product tweak, and asking for a new approach to managing them.

Also today: OpenAI's Decisions API is in public beta. It returns probabilities, picks or scores instead of text, on GPT-6 Luna, at ten cents per million input tokens.

TypeSafe AI closed $870 million at a $7.5 billion valuation, led by Andreessen Horowitz with Sequoia and DCVC, 24 days after launching Jev.

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