Daily AI Builder Brief - June 9, 2026
Executive Summary
- Claude Code 2.1.169 (June 8) adds
--safe-modeto disable all customizations for troubleshooting and/cdto switch directories without breaking the prompt cache. - 2.1.169 also fixes enterprise MCP policy enforcement on reconnect and stops background agents ignoring project-level env vars.
- 2.1.166 (June 6) introduced
fallbackModel(up to three models tried in order) and glob support in deny rules, so"*"now denies all tools. - Dynamic workflows in Claude Code went to research preview: Claude writes JavaScript orchestration scripts running tens to hundreds (up to 1,000) of parallel subagents in one session.
- The MCP 2026-07-28 release candidate is out: it removes the protocol-level session and
Mcp-Session-Idheader, so remote servers run behind a plain round-robin load balancer. - A fresh MCP CVE wave landed: VIPER-MCP swept ~40,000 repos for 67 CVEs, Akamai disclosed three database-MCP flaws (one unpatched), and the NSA published MCP hardening guidance.
Claude Code
2.1.169: safe-mode, /cd, and enterprise MCP fixes (June 8)
--safe-mode strips CLAUDE.md, plugins, skills, hooks, and MCP servers for clean troubleshooting. /cd changes working directory mid-session without a cache flush. Background agents now respect project env vars, and enterprise MCP policy enforces correctly on reconnect. Changelog
2.1.166: fallbackModel and glob deny rules (June 6)
Set up to three fallbackModel entries for automatic failover. Deny rules now take globs, so "*" blocks every tool. MAX_THINKING_TOKENS=0 disables thinking on default-thinking models. Changelog
Dynamic workflows hit research preview
Claude plans a task, writes an orchestration script, and fans work across parallel subagents with results checked before merge. Aimed at migrations, broad bug hunts, and security audits. Max, Team, and Enterprise plans. Announcement / Docs
Anthropic
Nothing new on the official newsroom inside the 48-hour window; latest post is June 3. The week's builder-relevant shift is the dynamic workflows preview above.
MCP Ecosystem
2026-07-28 release candidate: stateless core
The largest revision since launch. Drops the session layer, lets clients cache tools/list per ttlMs, and aligns auth with OAuth/OpenID Connect. Ten-week validation window before the final spec. MCP blog
Broader AI
New MCP vulnerability wave
VIPER-MCP produced 67 CVEs across ~40,000 server repos, Akamai disclosed three database-MCP flaws, and the NSA shipped design guidance. Roughly 40% of remote servers still expose tools with no auth. Adversa roundup
What This Means For Builders
- If you run remote MCP servers, start planning for the stateless RC now: you can drop sticky sessions and shared session stores, but you've got a ten-week window to validate.
- Audit your MCP servers this week. With 67 fresh CVEs and ~40% of remote servers unauthenticated, assume your dependencies are in scope.
- Use
fallbackModelto harden agent pipelines against single-model rate limits or outages. - Reach for dynamic workflows on migrations and audits that don't fit one context, but keep them on Max/Team/Enterprise and verify the merged output.
Sources
- https://code.claude.com/docs/en/changelog
- https://claude.com/blog/introducing-dynamic-workflows-in-claude-code
- https://code.claude.com/docs/en/workflows
- https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/
- https://adversa.ai/blog/top-mcp-security-resources-june-2026/
- https://www.anthropic.com/news