Daily AI Builder Brief - May 12, 2026

Executive Summary

Claude Code 2.1.139 shipped with an Agent view (Research Preview) and a /goal command that autoloops to a stated completion condition.
Hooks gain an args: string[] exec form, killing the shell-quoting bug that bit path placeholders.
MCP plumbing: new CLAUDE_PROJECT_DIR env var, better reconnect handling, fix for unbounded memory growth in MCP servers.
claude plugin details <name> now projects per-session token cost before you install.
Google attributed a zero-day exploited in the wild to an AI system, the first confirmed in-wild AI-built exploit.
Two Semantic Kernel CVEs (CVE-2026-25592, CVE-2026-26030) turn prompt injection into host RCE.
AWS MCP Server hit GA on May 6, still the freshest official MCP server worth wiring in.

Claude Code

v2.1.139 ships Agent view and /goal autoloop

Run claude agents for one list of every session: running, blocked on you, or done. /goal sets a completion condition and Claude keeps working across turns until it's met, with a live elapsed/turns/tokens overlay. Works in interactive, -p, and Remote Control modes. Changelog

Hooks get exec-form args, plugin cost previews land

args: string[] in hooks spawns the command without a shell, so path placeholders don't need quoting. continueOnBlock keeps PostToolUse hooks from halting on first block. claude plugin details <name> shows component inventory plus projected per-session token cost. Release notes

MCP plumbing fixes

CLAUDE_PROJECT_DIR now reaches MCP servers, reconnect logic improved, unbounded memory growth in MCP servers patched. Changelog

Anthropic

Nothing new in the last 24 to 48 hours. Most recent ship was May 6 at Code with Claude: doubled Claude Code rate limits, ten financial services agent templates, Claude add-ins for Microsoft 365, and dreaming for Managed Agents. Anthropic news

MCP Ecosystem

AWS MCP Server hit GA (May 6)

Single-tool access to any AWS API plus sandboxed Python script execution for multi-step ops. Still the freshest official server to wire into a builder stack. AWS announcement

Broader AI

Google confirms AI-built zero-day in the wild (May 11)

Google disclosed an unknown threat actor used a zero-day "likely developed with an AI system." First confirmed in-wild malicious use of AI for vuln discovery and exploit generation. The Hacker News

Semantic Kernel RCE (CVE-2026-25592, CVE-2026-26030)

One prompt is enough to launch calc.exe on the host running the agent. If you ship on Semantic Kernel, patch now. Microsoft Security Blog

What This Means For Builders

Wire /goal into long-running build/test loops where you can codify a pass condition, then let Claude grind.
If you publish Claude Code plugins, run claude plugin details first so users see honest token cost.
Audit any agent framework that interpolates LLM output into shell or template strings. The Semantic Kernel CVEs aren't an isolated pattern.
AI-built exploits aren't hypothetical anymore. Treat agent stacks as live attack surface today.