Anthropic Claude Code Security Updates
Announced at Code with Claude London, May 19-20, 2026 ================================================================ SELF-HOSTED SANDBOX (PUBLIC BETA) Claude Managed Agents can now run tool execution inside your own infrastructure instead of Anthropic-managed containers. Orchestration stays on Anthropic's side, but code, filesystem, and network egress stay in your environment. Works with Cloudflare, Daytona, Modal, or Vercel as managed providers. Built for orgs with data-residency, compliance, or internal-service- access requirements. ================================================================ SECURITY-GUIDANCE PLUGIN (FREE, ALL PLANS) Scans every file edit in real time for risky patterns like eval, innerHTML, command injection, and XSS. Acts as a pre-write validation hook that blocks unsafe writes. Powered by Opus 4.7 by default. Install via: /plugin install security-guidance Anthropic's internal testing showed 30-40% reduction in security- related PR comments. Catches roughly 25 dangerous patterns via regex matching, so it handles low-hanging fruit, not semantic analysis. ================================================================ CONTEXT Part of a broader security push. - Anthropic's Mythos model recently flagged 23,000+ potential vulnerabilities across 1,000+ OSS projects. - Enterprise integrations expanded to 28 security and compliance platforms. - Two Claude Code sandbox vulnerabilities (privilege escalation, network bypass) were patched in recent months, which explains the demand. ================================================================