Daily AI Builder Brief - May 4, 2026

Executive Summary

Claude Security public beta gets full press cycle on May 4, scanning whole repos with Opus 4.7 and pushing findings to Slack and Jira.
Claude Code 2.1.126 (May 1) lands gateway-aware /model picker, claude project purge, and 3-attempt MCP auto-retry on transient startup errors.
MCP 2026 roadmap names four expedited SEP areas: stateless transport, Tasks lifecycle, server discovery via .well-known, and enterprise readiness.
Google ships fully managed remote MCP servers across its API surface, giving builders a stable enterprise endpoint without self-hosting.
"Comment and Control" prompt injection disclosure from Johns Hopkins researchers hits Claude (CVSS 9.4), GitHub Copilot, and Google coding agents at once.
No fresh Anthropic news posts in the May 2-4 window. Today is integration day, not announcement day.

Claude Code

v2.1.126 keeps gateway parity moving

The /model picker now reads from the gateway's /v1/models endpoint when ANTHROPIC_BASE_URL points at an Anthropic-compatible gateway, so multi-tenant deployments stop showing stale lists. New claude project purge [path] wipes transcripts, tasks, file history, and config in one shot with --dry-run and -i modes. MCP servers that fail on startup auto-retry up to three times instead of dying silently. Vertex AI gains X.509 Workload Identity Federation. Changelog

Anthropic

Claude Security beta gets May 4 coverage push

Help Net Security and Infosecurity ran the deeper writeups today. The beta scans entire repos, directories, or branches, generates confidence ratings plus reproduction steps, and proposes patches, all on Opus 4.7. Webhook integrations with Slack and Jira are first-class. Partner ring includes CrowdStrike, Wiz, Palo Alto, and SentinelOne. Help Net Security

MCP Ecosystem

2026 roadmap names four priority lanes

Maintainers ranked stateless transport scaling, Tasks lifecycle (retry semantics, expiry policies), server discovery via .well-known Server Cards, and enterprise readiness (audit, SSO, gateway behavior) as expedited SEP review areas. Roadmap

Google ships managed remote MCP servers

Google announced fully managed remote MCP endpoints over its API infrastructure. Builders get a globally consistent endpoint for Google and Google Cloud services without running their own server. Google Cloud blog

Broader AI

"Comment and Control" hits three coding agents at once

Johns Hopkins researchers Aonan Guan, Zhengyu Liu, and Gavin Zhong disclosed a prompt-injection class affecting Claude, Copilot, and Google's coding agents. Anthropic rated it CVSS 9.4 Critical. GitHub paid via the Copilot Bounty. If your app embeds untrusted comments or PR descriptions into agent context, treat this as actionable today. VentureBeat

What This Means For Builders

If you run Claude through a gateway, upgrade to 2.1.126 so users see the right model list. Multi-tenant pain goes away.
Gate any agent that ingests user-controlled text (PR comments, issue bodies, doc strings) behind a stripping pass. Comment and Control is class-of-attack, not one-off.
Prefer Google's managed MCP endpoint over self-hosting a Google bridge. Less infra, same protocol.
The MCP roadmap signals that Server Cards become the discovery default. Start emitting .well-known metadata now.