Claude Report — 2026-09-24

🚀 Anthropic Official

Claude discovers a novel enzyme system with CRISPR-like repeats

Anthropic's biology research group has Claude autonomously identify an uncharacterized enzyme system — dubbed "array-associated reverse transcriptases" (ART) — hidden in bacteriophage DNA, structurally resembling CRISPR repeat arrays. The run used roughly 950 agents over 21 hours and consumed about 210 million tokens; the system's actual function remains unknown (Sep 23).

Claude Code v2.1.281 (Sep 23)

Ships on top of v2.1.280's Opus 5.5 default-model switch with a gateway-and-MCP-focused release: Claude Desktop gateway policy blocks (blockReadsOutsideWorkingDirectories, disableBypassPermissionsMode), Bedrock assume_role and guardrail config for gateway upstreams, telemetry.resource_attributes labeling, an "attribution": false settings option to hide commit/PR attribution, MCP URL-mode elicitation for browser-based auth flows, and new claude plugin validate checks that catch silently-dropped .mcp.json entries, undeclared ${user_config.*} references, and insecure URLs.

Official SDKs ship same-day parity with Claude Code v2.1.281 (Sep 23)

• claude-agent-sdk-typescript v0.3.281 — adds trigger, user_message_uuid, and timestamp fields to conversation_reset messages, fixes permission/dialog callbacks firing after close(), fixes control requests hanging after query closure, and shrinks sdk.mjs from 1.47MB to 0.97MB.
• claude-agent-sdk-python v0.2.159 — bundles CLI 2.1.281 and pins the e2e test default model to claude-opus-5 to work around CI failures caused by the new default.

More Opus 5.5 detail surfaces post-launch (Sep 22)

Benchmark disclosures show Opus 5.5 hitting 66.4% on Terminal-Bench 4.0 and 57.8% on CursorBench 4.0 (both above Fable 5.1), plus a previously unreported safety-routing detail: some cybersecurity requests can be transparently rerouted to Opus 4.8, and biology-safeguard triggers can similarly fall back to a weaker model.

🔌 Claude Code Plugins

Plugin tracking skipped this run — official marketplace install-count cache unavailable.

🛠️ Skills

None of the top 5 tracked skills updated in the window.

Unchanged in window: addyosmani/agent-skills (last commit Sep 23, already reported yesterday — no new activity), anthropics/skills (last commit Sep 10), cloudflare/security-audit-skill (last commit Sep 14), K-Dense-AI/scientific-agent-skills (no new commits since automated Sep 21 scan), eugeniughelbur/obsidian-second-brain, affaan-m/ECC, agentskillexchange/skills, tt-a1i/archify, ayghri/i-have-adhd (no substantive commits observed in window).

🤖 Agents & Subagents

nicobailon/pi-subagents

Continues past its v0.71.0 release (Sep 23) with new Sep 24 fixes: stops async TUI widget flicker while the Fleet roster panel is open, loads the council protocol through a guide topic, adds terminal-proof waiting for nested-control tests, and follows up child-launched external-job runs.

Unchanged in window: VoltAgent/awesome-claude-code-subagents (last commit Sep 21, README only), anthropics/commerce-agents (no commits since Aug 31).

🔗 MCPs & Integrations

oraios/serena

Fixes a PyJWT authentication-bypass CVE (HMAC key-confusion, CVE-2026-48526) by bumping to PyJWT 2.13.0, adds Astro language-server support, fixes AL language-server lookup on newer extension layouts, and drains in-flight TypeScript indexing on cross-file queries (Sep 23–24).

ChromeDevTools/chrome-devtools-mcp

Ships v1.10.1 (Sep 23) and follows up with better chrome-devtools CLI config support and updated docs/SKILL content for the get_css_styles tool (Sep 24).

upstash/context7

Adds a Search API documentation page and flattens the Docs7 sidebar, plus a new Docs7 custom-domains page (Sep 23–24), building on last window's Search API SDK launch.

DeusData/codebase-memory-mcp

Fixes a Windows endpoint directory-creation race condition, corrects C pointer return-type handling, fixes a daemon runtime sanitizer timeout, and switches the CLI's top-level --help to JSON format (Sep 23–24).

Unchanged in window: github/github-mcp-server (last commit Sep 16), modelcontextprotocol/servers (only a minor "everything" example-server subscription-cleanup fix Sep 22, no further activity).

💡 Community — Workflows & Ideas

davila7/claude-code-templates

Enhances three catalog agents in quick succession — ai-ethics-advisor (#949, Sep 21), accessibility-tester (#960, Sep 22), and ux-researcher (#976, Sep 23) — alongside a new jev-guardrails mod porting TypeSafe's guardrails cookbook and a --skip-downloads catalog regeneration mode.

📰 Quick Mentions

• Claude Opus 5.5 Intelligence, Performance and Price Analysis — artificialanalysis.ai benchmark thread (330 points) finds "medium" reasoning effort often outperforms "max," which can burn its full 128k budget without producing output.
• Claude Opus 5.5 — top HN launch thread (1,560+ points) debates Anthropic's "pacing the frontier" framing against the release's actual cadence.
• Claude Opus 5 Helped Researchers Take Over OpenAI Staff Accounts via Chained Flaws — security writeup on a red-team exercise using Claude to chain exploits against OpenAI staff accounts.