Daily AI Builder Brief - June 6, 2026

Executive Summary

Claude Code 2.1.166 shipped with fallbackModel (up to three fallbacks tried in order), now applied to interactive sessions too.
Deny rules now accept glob patterns in the tool-name slot; "*" denies all tools, and unknown deny-rule names warn at startup.
Security fix: messages relayed via SendMessage from other Claude sessions no longer carry user authority.
MAX_THINKING_TOKENS=0 and --thinking disabled now turn off thinking on models that think by default.
Anthropic Institute published "When AI Builds Itself," calling for a verifiable, multi-lab coordinated pause before recursive self-improvement.
Claude reportedly writes over 80% of its own codebase contributions, up from under 10% early last year.
A June 5 outage hit claude.ai, the API, Claude Code, and Cowork; full recovery confirmed 18:27 UTC.
No MCP protocol release inside the 48-hour window. The 2026-07-28 spec RC remains the thing to track.

Claude Code

2.1.166: fallbackModel and interactive fallback

Configure up to three fallbacks tried in order when the primary is overloaded. --fallback-model now works in interactive sessions, and a turn retries once on the fallback after an unexpected non-retryable API error. changelog

2.1.166: glob support in deny rules

Deny rules take globs in the tool-name position. "*" blocks everything, allow rules reject non-MCP globs, and unknown deny names warn at startup. changelog

2.1.166: cross-session messaging hardened

Relayed SendMessage payloads from other sessions no longer inherit user authority, closing a privilege path between sessions. changelog

Anthropic

"When AI Builds Itself"

Jack Clark and Marina Favaro argue labs are nearing full recursive self-improvement and propose a coordinated, cross-country slowdown each lab can verify in the others. Not a unilateral stop: a framework. report

June 5 platform outage

Elevated error rates across all models and Code/Cowork, resolved by 18:27 UTC. If your pipeline calls the API, this is the case fallbackModel and retries are built for. cybernews

MCP Ecosystem

No release in window

Nothing landed in the last 48 hours. The standing item is the 2026-07-28 spec RC (stateless core, Mcp-Method/Mcp-Name routing, ttlMs caching, Tasks and Apps extensions). Validate servers against it now. RC post

Broader AI

Nothing today.

What This Means For Builders

Set fallbackModel today: yesterday's outage is the exact failure it covers, and it now applies to interactive sessions.
Audit deny rules. Globs make blast-radius control real, but "*" plus a typo'd allow rule can lock you out: test before committing.
If you run agent swarms, the SendMessage authority fix matters: re-check any flow that assumed relayed messages were trusted.