Gather — 10 April 2026

Security Incidents (April 2026) #

• Claude Code Permission Bypass — Adversa AI disclosure (PiunikaWeb, 6 Apr 2026) — Adversa AI found that Claude Code’s 50-subcommand security-analysis limit creates a bypass: embed malicious commands after command #51 in a chain and deny rules no longer apply. Curl embedded in a long chain ships SSH/cloud credentials to attacker. Patched by Anthropic 6 Apr 2026.
• Anthropic Patches Claude Code Bypass Vulnerability (Let’s Data Science) — Confirms patch date and describes command-parser bug that silently bypassed developer-configured deny rules.
• Claude Code Vulnerability Exposes New AI Security Risks (Seceon) — Security-vendor analysis framing the permission-bypass as a class of “agent trust boundary” vulnerabilities distinct from prompt injection.
• Claude Code News April 2026 — Startup Edition (Mean CEO) — Aggregator summarising the month’s incidents; useful timeline.

Security Research by Claude Code #

• Claude helps researcher dig up decade-old Apache ActiveMQ RCE vulnerability (CVE-2026-34197) (Help Net Security, 9 Apr 2026) — Claude Code found an RCE vulnerability that had been hidden in Apache ActiveMQ for a decade. Concrete, recent, high-signal use case.
• Claude Code found a Linux vulnerability hidden for 23 years (Adafruit Blog, 7 Apr 2026) — 23-year-old Linux kernel vulnerability discovered via Claude Code. Strong signal for Claude’s code-comprehension reach.
• Security researchers frustrated as Claude Code rejects vulnerability research (PiunikaWeb) — Tension: Anthropic’s safety filters increasingly block legitimate security research. Researchers report Claude refusing to analyse obviously vulnerable code patterns.

Simon Willison (April 2026) #

• Tool: Cleanup Claude Code Paste (Simon Willison, 6 Apr 2026) — New utility for cleaning up code pasted from Claude Code sessions. Small but canonical workflow friction point.
• claude-code-transcripts — CLI tool for HTML transcript extraction (GitHub) — Publishes both local and Claude-Code-for-web session transcripts as readable HTML. Useful for auditable records of agent runs.
• A new way to extract detailed transcripts from Claude Code (Substack) — Blog-post companion to the CLI tool; walkthrough.
• Claude Skills are awesome, maybe a bigger deal than MCP (Substack) — Willison’s thesis: Skills (self-contained SKILL.md packages) are a more important primitive than MCP for most practical agent work. Contrarian but from a trusted source.

Workflow + Plugin Ecosystem #

• Claude Code Skills vs MCP vs Plugins: Complete Guide 2026 (MorphLLM) — Disambiguates the three primitives. Noise-filter concern: commercial source.
• 10 Must-Have Skills for Claude (and Any Coding Agent) in 2026 (Medium) — Curation of battle-tested skills; agent-skill standard spreading beyond Claude Code to Codex, Gemini CLI.
• Anthropic: Create plugins — Official Docs (Claude Code Docs) — Canonical plugin-authoring reference.
• alirezarezvani/claude-skills — 220+ skills for 9 coding agents (GitHub) — Largest cross-agent skills collection; demonstrates the Agent Skills standard crossing Claude Code boundaries.
• jeremylongshore/claude-code-plugins-plus-skills — 340 plugins + 1367 skills (GitHub) — Scale indicator: the ecosystem is now measured in thousands of skills, not dozens.
• Claude Code Marketplaces aggregator (claudemarketplaces.com) — Community marketplace index.
• Claude Code Hooks Complete Guide (March 2026 Edition) (SmartScope) — Updated hooks reference covering the 21 lifecycle events in the current docs.
• Claude Code hooks: A practical guide with examples (Eesel AI) — /hooks debug command, testing hooks via piped stdin, event-logger.py for discovering event payloads.

Advanced Workflows (Frontend Masters + Community) #

• Claude Code Deep Dive — Frontend Masters Workshop (Apr 21, 2026) (Frontend Masters) — Dedicated advanced workshop: harness-vs-model mental model, MCP server authoring, Skills encoding, signal-tracking for setup quality.
• Claude Code Best Practices — Official Docs (Anthropic) — Canonical reference; has been meaningfully updated since last gather.

Cross-links #

• Applications of Vibe Coding “Agent trust boundary” class of vulnerability (permission bypass) is the enterprise-governance counterpart to the comprehension-debt discourse.
• Vibe Coding Approaches Skills-vs-MCP-vs-plugins disambiguation is the primitive-layer debate underneath agentic-engineering methodology.
• AI Impact on Society Claude Code finding 23-year-old vulnerabilities is a concrete “augmentation not replacement” data point for the workforce-transformation narrative.
• Open vs Closed AI Ecosystems The Agent Skills standard crossing Claude Code → Codex → Gemini CLI is a rare convergence signal across closed-lab silos.

Meta-observations #

• Emerging theme: Claude Code’s security story has two halves that are in tension — it’s finding decades-old vulnerabilities in ActiveMQ/Linux while shipping new ones (permission bypass, source leak). The net trust calculus is unclear. Worth tracking as a paired metric.
• Emerging pattern: The Skills primitive is gaining momentum (Willison: “maybe bigger than MCP”; 220+ and 1367+ skill collections). If the Agent Skills standard generalises to Codex/Gemini CLI, this is a cross-lab protocol win — and a Claude ecosystem leadership signal.
• Quality signal: Adversa AI has now published two consecutive high-quality Claude Code vulnerability disclosures. Promote to source-to-watch.
• Source to watch: Adversa AI — adversarial-research firm producing rigorous Claude-specific security work.
• Source to watch: Help Net Security — publishing recent, dated security-research reporting on AI agents.
• Keyword suggestion: “claude code permission bypass” / “agent trust boundary” — new class of vulnerabilities worth tracking beyond prompt injection.
• Keyword suggestion: “claude code vulnerability research” — the “Claude finding bugs” story is a major use-case distinct from “Claude having bugs.”
• Gap: Still missing substantive Boris Cherny / Anthropic-team content since the last gather. Possible the flow has slowed post-launch, or the search query needs sharpening.
• Noise pattern: “Claude Code best practices 2026” listicles are multiplying. The exclude_terms filter is doing its job but marketplace-vendor blogs (morphllm, eesel, serenitiesai) are filling the gap — may need selective exclusion.