Brex built its AI agent policy by watching what agents actually do, not by writing rules first
THE SO WHAT
Brex learning that traditional guardrails couldn’t safely contain agents with real API keys and service accounts is the operational reality check for agentic AI. If you’re piloting agents, instrument behavior first and derive policy from observed failure modes instead of starting with a static rules doc.
READ THE SOURCE
MORE FROM THE WIRE
Applied AITikTok is testing an AI likeness detection tool
AI likeness detection moving into TikTok’s workflow means platforms are starting to operationalize deepfake governance at scale — but only for creators who opt in. Brands and public figures should assume fragmented protection for now and maintain their own monitoring for impersonations across platforms.
Applied AISources: SpaceX is in talks with the DOD about providing the agency with access to data-center capacity worth billions of dollars for running AI models
The Pentagon contemplating billions in AI data-center capacity from SpaceX shows defense AI is moving from pilots to industrial-scale compute procurement. Contractors should expect future RFPs to assume access to hyperscale or near-hyperscale GPU clusters — and align their architectures and partners accordingly.
Applied AIAgents think in milliseconds, legacy infrastructure doesn't. LinkedIn, Walmart and Zendesk shared how they closed the gap at VB Transform 2026
If LinkedIn, Walmart, and Zendesk all say legacy infra is the real bottleneck for AI agents, the constraint has shifted from model quality to system latency and orchestration. Before you chase the next model upgrade, profile your end-to-end request path — network hops, DB calls, service fan-out — and remove the 100–500 ms anchors.
Applied AIWhat’s next in Apple’s legal battle with OpenAI
Apple’s lawsuit against OpenAI hanging over potential hardware ambitions injects legal risk directly into AI device roadmaps. If you’re planning tight OS–model integration, you now have to underwrite IP and partnership risk alongside technical feasibility.