Claude Autonomously Discovers a Novel CRISPR-Like Enzyme System — 950 Agents, 21 Hours, and Scientists Still Aren't Sure What It Does

950 agents, 21 hours — Claude found an enzyme system that structurally resembles CRISPR, but Anthropic itself admits it doesn't yet know what the system actually does.

Ben Calloway · September 28, 2026

On September 23, 2026, Anthropic announced that its newly established San Francisco biology lab, running an autonomous research task on Claude, identified a previously undocumented enzyme system — named Array-Associated Reverse Transcriptases (ART) — with structural features that reminded researchers of CRISPR gene-editing systems. This is one of the first publicly documented cases of an AI agent being credited with independently completing the full search process behind a concrete...

practice

Claude Code Adds Model Version Locking: How availableModelsMatch: "exact" and deniedModels Actually Work

availableModelsMatch: "exact" plus deniedModels upgrades model version...

practice

Claude Code's New rm Command Rule: A Two-Minute Timeout That Denies Instead of Hanging Forever

Confirmation on a dangerous command no longer hangs forever — a two-minute...

beginners

How to Set Up Claude Code's Screen Reader Mode: A Complete Guide for Blind and Low-Vision Developers

This mode doesn't just change the display — even the timing is redesigned...

Useful Resources

Claude API Status → Model Pricing → Prompt Playground → Token Counter → MCP Servers → LLM Benchmarks → Model Comparison →

Glossary

View All →

Agent Permission Scoping

The practice of deliberately granting an <a href="/en/glossary/core-concepts/ai-agent/">AI Agent</a> only the minimum set of permissions needed for its current task, rather than granting broad access far beyond actual need for convenience, so the potential impact of an error stays as contained as possible.

What is <a href="/en/glossary/agent-permissions/agent-permission-scoping/">Agent Permission Scoping</a>, and how does it differ from the intuition...

advanced

Least-Privilege Agent

Granting an agent only the minimum set of tools and data access strictly needed for its current task, rather than starting broad and narrowing later.

What does the principle of least privilege actually mean when applied to an agent? Least privilege isn't an AI-specific concept — it's a...

intermediate

Prompt Injection Defense

Designing a mechanism so an agent, while processing external data (like webpage content or user-uploaded documents), can distinguish original task instructions from malicious instructions potentially hidden within the data itself, avoiding being misled into performing unintended actions by content buried in that data.

What is prompt injection defense, and how does it differ from typical <a href="/en/glossary/prompt-techniques/jailbreaking/">jailbreaking</a>...

intermediate

Weekly Picks

practice

Claude Code's Auto Mode Classifier Is Now Free — Unless You're Behind a Gateway, Where It Quietly Isn't

Auto Mode's safety classifier is free now — unless you're behind a gateway, where you're still...

Marcus Ide

practice

Claude Code Now Reads AGENTS.md: The Four instructionFiles Modes and a CLAUDE.local.md Priority Trap

CLAUDE.md and AGENTS.md aren't a simple override — it's a per-directory either/or, and a...

Derek Finch

practice

How Claude Code Projects' Shared Memory Actually Works: One MEMORY.md Decides What Every Parallel Thread Knows

Memory is shared across the whole project. Permission rules only apply within the launch...

Derek Finch

practice

Claude Code Projects Caps at 200 Threads a Day, But the Real Bottleneck May Not Be That Number

Hitting the 200-thread cap just means waiting. What actually leaves you gasping is 10 parallel...

Cora Mitchell