Daily AI Builder Brief - June 8, 2026

Executive Summary

Claude Code 2.1.166 adds fallbackModel (up to three models tried in order when the primary is overloaded).
Same release adds glob patterns in MCP deny rules: "*" denies all tools in one line.
Cross-session messaging hardened: messages from other Claude sessions no longer carry your user authority.
Thinking is now fully disable-able (MAX_THINKING_TOKENS=0, --thinking disabled) even on models that think by default.
Anthropic mapped a year of AI-enabled attacks to MITRE ATT&CK: medium-or-higher risk actors jumped from 33% to 56%.
The differentiator isn't technique count, it's the autonomous "scaffolding" attackers build to chain steps.
Strava shipped an official remote MCP connector, Claude-first.

Claude Code

2.1.166 to 2.1.168 (June 5 to 6)

Feature work landed in 2.1.166: fallbackModel setting and --fallback-model for interactive sessions, MCP deny-rule glob patterns, hardened cross-session permission relays, per-model thinking toggle, a one-retry-on-fallback path for unexpected API errors, JetBrains terminal flicker fix, and a Kitty-protocol Shift+non-ASCII fix. 2.1.167 and 2.1.168 were reliability follow-ups. Changelog

Anthropic

Mapping a year of AI-enabled cyber threats (June 3 to 4)

Anthropic analyzed 832 banned accounts against MITRE ATT&CK. AI use is shifting from initial access toward post-compromise activity (account discovery up 8.9%), and 67.3% used AI to write malware. Report

MCP Ecosystem

Toward a stateless core

The next MCP spec drops the protocol-level Mcp-Session-Id, so any request can land on any instance: no sticky sessions, plain round-robin load balancing, cacheable tools/list. MCP Blog

Strava official connector (June 1)

Strava began rolling out a remote MCP server for subscribers, OAuth-based, launching Claude-first. Strava

Broader AI

Nothing today.

What This Means For Builders

Set fallbackModel now: overload no longer stalls an agent run, it degrades gracefully.
If you ship MCP tooling, audit deny rules with the new globs and assume cross-session messages are untrusted.
The ATT&CK data says defenders should model autonomous chaining, not lone techniques: test your own agents for the same scaffolding.
Design remote MCP servers stateless now so they're ready for the round-robin spec.