Back to AI Briefs
TLP:CLEARAI-2026-0601

Daily AI Builder Brief - June 1, 2026

June 1, 20263 min read
aiclaude-codeanthropicmcpbuilder-intel

Executive Summary

  • Claude Code 2.1.158 brings Auto mode to Bedrock, Vertex, and Foundry for Opus 4.7 and 4.8 (opt in with CLAUDE_CODE_ENABLE_AUTO_MODE=1).
  • 2.1.159 (May 31) was internal infra only: nothing user-facing to act on.
  • Plugin auto-loading from .claude/skills and claude plugin init scaffolding (2.1.157) change how you package and ship agent tooling.
  • A joint OpenAI, Anthropic, and Google DeepMind paper resurfaced: every tested AI defense broke above 90 percent under adaptive attack.
  • The MCP stateless-core release candidate is locked, dropping the session header so servers run behind plain round-robin load balancers.
  • Opus 4.8 is now the cloud-provider Auto mode target, not just the API default.

Claude Code

Auto mode on Bedrock, Vertex, and Foundry (2.1.158)

Auto mode now runs on all three managed-cloud backends for Opus 4.7 and 4.8. Set CLAUDE_CODE_ENABLE_AUTO_MODE=1 to opt in. changelog

Plugin auto-loading and scaffolding (2.1.157)

Plugins in .claude/skills load without a marketplace, and claude plugin init <name> scaffolds new ones. The agent field in settings.json is now honored for dispatched sessions. changelog

Anthropic

Opus 4.8 in production rotation

The May 28 Opus 4.8 release is now the model behind cloud Auto mode, and 2.1.156 fixed a thinking-block bug that was throwing API errors on 4.8. If you pinned 4.7 to dodge that, you can re-test. anthropic.com/news

MCP Ecosystem

Stateless-core release candidate locked

The RC removes Mcp-Session-Id and the protocol-level session, so any request lands on any instance. Remote servers drop sticky routing and shared session stores, running behind round-robin load balancers with client-cached tools/list. Final spec lands July 28. MCP RC

Broader AI

Every tested AI defense bypassed above 90 percent

"The Attacker Moves Second" (Carlini, Nasr, Hayes, Shumailov, Tramer) re-tested 12 published defenses with adaptive attacks: 95 to 99 percent success on prompting defenses, 96 to 100 percent on training-based ones. Static-attack benchmarks overstate safety. VentureBeat / arXiv

What This Means For Builders

  • If you run agents on Bedrock, Vertex, or Foundry, you can finally turn on Auto mode for Opus 4.8 instead of self-hosting orchestration.
  • Move shared agent tooling into .claude/skills plugins now: auto-loading kills the marketplace step in your CI.
  • Plan MCP servers for statelessness today so the July 28 spec is a config change, not a rewrite.
  • Treat any "near-zero attack success" claim from a security vendor as unproven until it survives adaptive red teaming. Ask for the adaptive results.

Sources

  1. https://code.claude.com/docs/en/changelog
  2. https://www.anthropic.com/news
  3. https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/
  4. https://blog.modelcontextprotocol.io/posts/2026-mcp-roadmap/
  5. https://venturebeat.com/security/12-ai-defenses-claimed-near-zero-attack-success-researchers-broke-all-of-them
  6. https://arxiv.org/abs/2510.09023