Claude Report — 2026-10-18

🚀 Anthropic Official

Claude reveals ECDSA curve origins (Oct 16)

Anthropic's Frontier Red Team reports that an internal Claude research model reconstructed, in under 30 minutes, how the NSA generated the seeds for NIST's P-192/224/256/384/521 curves: outputs of the C library's lrand48 generator seeded with Unix timestamps from 1997-1998, with no hash algorithm involved. Claude recognized the pattern and solved for the generator's internal state using LLL lattice-basis reduction; cryptographer Filippo Valsorda independently verified the result, and the $36,864 bounty went to the Internet Security Research Group (Let's Encrypt). The finding indicates the NSA's choice was conventional rather than a deliberate backdoor.

How the Claude Code team does code review (Oct 15)

Anthropic engineers detail restructuring code review around agent-generated volume (2,500 to 30,000+ changes/day): pre-PR self-verification, AI review on every PR that flags bugs before humans see them, a security-review gate that blocks merges, and risk-based human approval scaled by change type. They built feedback loops using incident data as eval sets, added convergence limits to stop infinite agent-reviewer loops, and cut reverts by 40% while reducing required human approvals.

🔌 Claude Code Plugins

Plugin tracking skipped this run — marketplace install-count cache unavailable, per instructions.

🛠️ Skills

None of the top 5 updated in window.

Unchanged in window: addyosmani/agent-skills (last commit Oct 13), anthropics/skills (last commit Oct 9), kharmanskyi/open-steps (last commit Oct 5), K-Dense-AI/scientific-agent-skills (last commit Oct 12), cloudflare/security-audit-skill (last commit Sep 14).

🤖 Agents & Subagents

wshobson/agents (Oct 15-17)

Ships a run of fixes and one new skill across this 185-skill collection: adds social-evidence-research for traceable social-discussion analysis, repairs broken examples in three skills (bash defensive patterns, async Python, auth middleware), corrects Cypress/Playwright API usage in e2e-testing-patterns, and migrates monorepo-management guidance to Turborepo 2 and pnpm workspaces.

nicobailon/pi-subagents v0.79.1 (Oct 18)

Cuts per-request overhead by withholding the bg_wait and subagent_supervisor tool definitions (~1,500 characters) until a session actually delegates to a subagent, fixes embedded hosts (like pi-web) launching children with the wrong bundled Pi install, and preserves already-completed child results when a workflow's main process dies.

davila7/claude-code-templates (Oct 18)

Adds a bluesky-feed mod that surfaces a Bluesky timeline pane inside Claude Code, on top of this week's NVIDIA SkillSpector scoring badge added to every catalog component.

Unchanged in window: VoltAgent/awesome-claude-code-subagents (last commit Oct 11), anthropics/commerce-agents (last commit Aug 31).

🔗 MCPs & Integrations

ChromeDevTools/chrome-devtools-mcp v1.11.0 (Oct 15)

Adds a --file-navigations flag to reject file: URL navigations, a new analyze_heapsnapshot_contexts tool, an ignoreCache option for performance_start_trace, and pagination plus name-truncation controls for compare_heapsnapshots.

upstash/context7 ctx7@0.8.0 (Oct 16)

Removes the deprecated skills and generate commands from the ctx7 CLI; ctx7 setup continues to install the find-docs and context7-mcp skills.

DeusData/codebase-memory-mcp v0.12.1 (Oct 16)

Fixes v0.12.0's npm/PyPI/Go installs, which all refused to unpack the new 331-335MB binaries due to a stale 256MB archive-member limit in each wrapper; also updates the stale Homebrew formula to ship a portable Linux static build and fixes a broken pinned go install.

oraios/serena (Oct 15)

Fixes symlink handling in request_full_symbol_tree and adds explicit file-path validation to Serena's LSP API, rejecting links to outside-root targets in untrusted projects while still allowing them in trusted ones — closing a path-traversal-adjacent gap flagged in issue #2163.

Unchanged in window: modelcontextprotocol/servers (last commit/release Oct 11).

💡 Community — Workflows & Ideas

Detecting Filesystem Capability Laundering in MCP Servers at Scale (Oct 15)

Researchers scan 21,000+ MCP servers and find 787 of 4,806 triaged servers (16.4%) expose filesystem capabilities despite being classified as non-filesystem tools, confirming 150 exploitable instances via dynamic testing; coordinated disclosure produced 13 advisories, including two CVEs in Anthropic's own Git MCP Server.

Vibemetric (Oct 17)

A macOS app scores Claude Code and Codex sessions across ten dimensions — testing habits, cost efficiency, security, planning, memory management — and recommends concrete workflow fixes.

Epoptes (Oct 18)

Lets a user describe a goal once in chat while the tool runs long, autonomous Claude Code or Codex sessions across multiple context-refresh cycles with live monitoring and steering.

claude-env-tabs (Oct 18)

Adds a pinned-projects pane with live session-state dots and a persistent tab bar to Claude Desktop's Code tab, for juggling multiple project sessions at once.

📰 Quick Mentions

• 500 clocks in one 6.3MB HTML file (Oct 18) — a developer's Opus-5.5-generated clock collection, built as a personal "pelican-on-a-bicycle"-style capability benchmark.
• Astrax (Oct 17) — new open-source MCP server giving agents see/click/type control over a Windows desktop.
• MCPX (Oct 18) — new MCP server directory that scans listings for tool poisoning and leaked keys.