Claude Code Daily Briefing - 2026-05-20

Release Summary

VersionDateKey Changes
v2.1.1445/18Exact Edit string matching, /resume shows background sessions, zip/URL plugin loading, Ctrl+R cross-project history
v2.1.1435/15Plugin dependency management, marketplace token cost display, worktree.bgIsolation: "none"

Full Release Notes

New Features & Practical Usage

v2.1.144 — Exact Edit Matching, Background Session Integration, Plugin Loading Expansion (5/18)

v2.1.144 significantly improves tool reliability and session management.

Exact String Replacement in Edit Tool: The Edit tool now performs exact string replacements, preventing unintended partial matches. This dramatically improves reliability when editing large files.

/resume Shows Background Sessions: /resume now lists background sessions (started via --bg or Agent View) alongside interactive ones, tagged with bg for easy identification.

/model Is Session-Only by Default: /model now only affects the current session. Press d in the model picker to set a default for all new sessions. This means experimental model switches won’t bleed into other sessions.

Zip/URL Plugin Loading: Plugins can now be loaded directly from zip files or URLs — useful for distributing internal plugins without a registry.

Ctrl+R Cross-Project History Search: Ctrl+R now searches history across all projects, not just the current one.

Hard Deny Rules for Auto Mode: You can now configure deny rules that block specific commands from ever executing in auto mode, structurally preventing dangerous commands from being auto-approved.

75s Startup Hang Fixed: Startup no longer hangs up to 75 seconds when api.anthropic.com is unreachable (captive portals, firewalls, VPNs) — side-channel API calls now timeout correctly at 15 seconds.

# /model session-only vs default
# Current session only: /model → select model
# All new sessions: /model → select model → press d

GitHub v2.1.144 | Changelog

Code with Claude London Extended — Today (5/20)

Following the Code with Claude London main event (5/19), the Extended session runs today. It targets indie developers and early-stage startups with hands-on workshops and 1:1 office hours. The main event featured three parallel tracks: Research, Claude Platform, and Claude Code. Recordings will be published later.

Next up: Tokyo (6/10) main event and Tokyo Extended (6/11).

London Extended | Tokyo Registration

Developer Workflow Tips

/model Session-Only Switch — Model Strategy Made Easy

With v2.1.144 making /model session-only, it’s now natural to use different models per session. Use Opus for architecture design, Sonnet for boilerplate generation, and Haiku for documentation — while keeping your default set to your most-used model.

# Set default model to Opus (one-time)
# /model → select opus → press d

# Use Sonnet for this session only (won't affect others)
# /model → select sonnet

The key principle: “defaults stay safe, sessions stay flexible.” You can safely A/B test models without affecting other sessions.

Claude Code Changelog

Hard Deny Rules — Structurally Block Dangerous Commands in Auto Mode

When running Claude Code in auto mode (--dangerously-skip-permissions or auto-accept), hard deny rules in v2.1.144 provide a structural safety net. Specific command patterns are blocked at the execution layer — even if Claude proposes them, they simply won’t run.

This is especially critical for CI/CD pipelines and unattended agent execution. Combine with Hook exit code 2 (block) for defense in depth.

Claude Code Docs

Security & Limitations

TrustFall — One-Click MCP RCE Across 4 AI Coding Tools (5/7)

Security firm Adversa AI disclosed TrustFall, a supply-chain vulnerability affecting Claude Code, Gemini CLI, Cursor CLI, and GitHub Copilot CLI simultaneously. Opening a malicious repository and pressing Enter once is enough for arbitrary code execution.

Attack mechanism: A malicious repo includes MCP server configuration. When a user approves the trust prompt, the MCP server launches as a native OS process with full user privileges — no sandbox, no directory restriction, no filesystem confinement.

Impact: Reading SSH keys, tokens, and secrets; accessing other projects; installing backdoors; establishing C2 connections.

Recommendations:

Dark Reading | Help Net Security | Adversa AI

Claude Code OAuth Token Theft via MCP Hijacking (5/16)

Security firm Mitiga demonstrated how Claude Code OAuth tokens can be stolen through MCP proxy hijacking. A malicious npm package’s lifecycle hook modifies the MCP server configuration in ~/.claude.json to point to an attacker’s proxy. Every time Claude Code initiates an MCP session, the token is forwarded to the attacker.

Persistence is key: Token rotation doesn’t help — the hook rewrites the token on the next load. Fixing the MCP URL is also reverted by the hook.

Anthropic’s response: Reported on 4/10, Anthropic replied on 4/12 that it was “out of scope,” stating the user had already consented to the behavior.

Recommendations:

SecurityWeek | Mitiga | eSecurity Planet

Ecosystem & Plugins

Anthropic Pursuing $900B Valuation with $30B Fundraise (5/12)

Anthropic is in talks to raise $30 billion at a valuation exceeding $900 billion. Greenoaks, Sequoia, Altimeter, and Dragoneer are co-leading, each investing $2B+.

Context: This is a 2.4x jump from the $380B Series G valuation just 3 months ago in February. The company has reached $30B in annualized revenue. If finalized, Anthropic would surpass OpenAI ($850B) as the world’s most valuable private company.

For developers, this signals continued capacity expansion for Claude Code — more compute means higher limits and better availability.

Bloomberg | TechCrunch

Google I/O 2026 — Gemini 3.5, Antigravity CLI Debut (5/19–20)

Google I/O 2026 introduced direct competitors to Claude Code’s territory:

Antigravity CLI is a direct Claude Code competitor, and its built-in terminal sandboxing is notably Google’s answer to security issues like TrustFall.

Google Developers Blog | 9to5Google

Community News

Minor Changes

Interesting Projects & Tools

By subscribing, you agree to our Privacy Policy.