Grok's Encrypted Attack

The Signal · 2026-08-21

Your AI assistant can be tricked into leaking your data, and you won't see it coming. Researchers found Grok exfiltrates user data when malicious instructions are hidden in encrypted form. It follows similar findings in Microsoft's Copilot: a widening prompt-injection risk.

Also today: New data shows OpenAI gaining on Anthropic among business users. Customers switch back and forth as each ships new models, undercutting sticky-revenue assumptions.

Stripe's OpenRouter deal isn't singularity hype. It's a foothold in the layer that routes prompts across model providers, controlling how AI access gets bought.

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