Format
Trust & Security
SB 53's disclosure mandate doesn't cover evaluation-related containment failures. The Hugging Face breach — and the guardrail paradox that followed — reveals how far the law trails the capability.
◆ Heath Callahan
Money & Markets
The stablecoin yield debate has split Wall Street along a structural fault line — investment banks want market structure, commercial banks want to protect $6 trillion in deposits. Section 404 is where the fight lives.
◆ Nolan Pratt
Infrastructure
The July 28 finalization cements protocol mechanics—stateless sessions, Extensions framework, 12-month deprecation windows—while Google Cloud forces migration to its Agent Registry two days later, creating a de facto standard that may diverge from the official spec.
◆ Blair Hayes
Trust & Security
The July 28 revision removes session management from the Model Context Protocol, forcing developers to implement security that the ecosystem is already failing to provide.
◆ Heath Callahan
Commerce
While consumer AI commerce remains stalled by trust deficits, the B2B sector is quietly moving from protocol signaling to actual, automated procurement.
◆ Tessa Vaughn
Infrastructure
Google Cloud is terminating its legacy Cloud API Registry API on July 30, forcing all MCP management workflows to the new Agent Registry—two days after the MCP Stateless Spec finalizes.
◆ Blair Hayes
Models
The AI inference startup doubled its valuation in seven months without shipping hardware. The money is the message.
◆ Lena Park
Trust & Security
CVE-2026-6875 lets unauthenticated attackers escape ServiceNow's script sandbox and execute arbitrary code across 85% of the Fortune 500. Five days after patches shipped, it was already being exploited in the wild.
◆ Heath Callahan
Trust & Security
Anthropic's containment architecture for Claude Code and Claude Cowork reveals a hard truth: model-level safeguards are probabilistic and will fail. The defense that actually works is deterministic, environment-level, and boring.
◆ Heath Callahan
Models
A joint UK-US evaluation found the strongest open-weight model for cyber operations has no safeguards against offensive use. On July 27, anyone can download it.
◆ Lena Park
