Anthropic AI Updates: October 14, 2026

1. Anthropic Replaces Git With In-House Version Control System Scion

Anthropic. Anthropic described Scion, an in-house version control system it built and migrated its monorepo to from git in five months after merges grew from about 2,500 per day in May 2026 to more than 30,000 per day by September. The repository has about 100 GB of main-branch history and serves roughly 25 million CI checkouts a day, which Anthropic estimates would have meant over 600 PB of daily git traffic. Scion treats clients as always online: commits, merges, and rebases run server-side as changesets backed by Spanner and content-addressed blob storage, and clients fetch objects on demand through a per-node daemon, regional caches, and an authoritative server, each tier reporting cache hit rates above 99%. Anthropic reports a roughly 200x smaller CPU and memory footprint than its previous git mirror fleet, CI checkouts dropping from minutes to seconds, and about 25% more CI CPU capacity; Scion is not being open-sourced. Source

2. Claude Code 2.1.299 Adds Idle Compaction Ahead of Prompt Cache Expiry

Anthropic. Claude Code v2.1.299 adds idle compaction in the terminal, which compacts a conversation of 200K or more tokens left idle for about 54 minutes so that the next message does not reprocess the full context after the one-hour prompt cache expires. The threshold can be changed with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS or disabled with "idleCompaction": false in settings.json. The release also adds CLAUDE_CODE_DISABLE_FILE_MENTIONS so @ mentions pass file paths instead of contents, a --shutdown-drain-sec option for the self-hosted runner orchestrator, --disallow-tools for claude plugin eval, and new /usage warnings for max effort, workflows, always-on context, and large tool results. Source

3. Uber CISO Details AI Security Stack Built Around Enterprise Frontier Safeguards

Anthropic. A Claude blog interview with Uber CISO Philip Martin describes how Uber’s security team uses Claude Fable to gather threat-modeling context from code, design docs, and Jira history, and how Uber worked with Anthropic on its Enterprise Frontier Safeguards product. Uber combines an LLM gateway, its open-source AI Detection and Response (ADR) tool, which collects real-time telemetry and acts inside the agent harnesses Uber deploys, and output from Anthropic’s classifiers. Martin says Uber retains control of its data under the safeguards and decides how to respond when one triggers, and he argues companies need an insider threat program for AI models similar to the ones they run for employees. Source