Anthropic AI Updates: May 27, 2026

1. Anthropic opens a Korea office and names KiYoung Choi as Representative Director, citing Claude usage at 3.5x the population baseline

Anthropic. Anthropic appointed KiYoung Choi as Representative Director of Anthropic Korea ahead of opening a Seoul office, the company’s third APAC location after Tokyo and Singapore. Choi joins from a 30-year run through senior roles at Snowflake, Google Cloud, Adobe, Autodesk, and Microsoft, and will lead enterprise partnerships, developer engagement, and startup outreach from Seoul. Anthropic cited Korea as one of its most active markets, with Claude consumption running more than 3.5x the rate expected for the country’s population, concentrated in technical and creative use cases. Source

2. Engineering blog details three containment patterns for Claude across products: gVisor, OS sandboxes, and isolated VMs

Anthropic. Anthropic’s engineering team published a deep dive on how Claude is sandboxed differently in each product to cap agent blast radius. Claude.ai runs code in ephemeral gVisor containers on isolated infrastructure for minimum capability and minimum risk, suitable for short-lived analysis. Claude Code uses a human-in-the-loop OS-level sandbox with Seatbelt on macOS and bubblewrap on Linux, requiring user approvals for risky operations. Claude Cowork runs work inside virtual machines with credentials held in the host keychain and code executing in a sealed guest with granular file-mount controls, layered with model-level classifiers and system prompts. Source

3. Code w/ Claude London 2026 ships self-hosted sandboxes in public beta and MCP tunnels in research preview for Managed Agents

Anthropic. At Code w/ Claude London 2026, Anthropic announced a batch of updates to Claude Managed Agents centered on running agent tools inside customer-controlled infrastructure. Self-hosted sandboxes enter public beta, allowing tool execution on a customer’s own machines, and MCP tunnels enter research preview to extend Model Context Protocol over private network connections without exposing services publicly. Boris Cherny, Head of Claude Code, framed the moment by comparing it to “the calculator feeling, except the calculator can write a distributed system.” Spotify, Base44, and Legora presented production engineering workflows built on Claude Code. Source

4. HubSpot product and marketing leaders publish a Q&A on scaling Claude Cowork across non-engineering teams

Anthropic. Anthropic published a Q&A with HubSpot’s product and marketing leadership detailing how the company rolled out Claude Cowork beyond engineering. The piece covers rollout patterns, governance choices, and the team workflows that emerged after broader deployment, framed as enterprise adoption mechanics rather than a single use case. It appeared in tandem with a Twilio customer story published the same day, signaling a coordinated enterprise push for Cowork timed to the London event. Source

5. Twilio details a self-learning Claude Code developer platform that compounds agent interactions into reusable workflows

Anthropic. A Twilio product manager described an internal developer platform built on Claude Code that captures every agent interaction and feeds it back into subsequent runs, compounding institutional knowledge over time. The platform is positioned as a way to turn one-off agent sessions into a self-improving system rather than a stateless tool. Twilio’s piece was published the same day as the HubSpot Cowork Q&A, underlining Anthropic’s enterprise narrative around agentic platforms inside large customers. Source

6. Anthropic runs a public webinar showing how its own sales leader plans the week with Claude Cowork

Anthropic. Anthropic hosted a webinar walking through how its own sales leadership uses Claude Cowork day-to-day to plan, prioritize, and execute against the week. The session positions Cowork as a productivity surface for non-engineering teams, with Anthropic visibly eating its own dog food. The webinar pairs with the same-day HubSpot and Twilio customer stories as part of a coordinated content push for enterprise Cowork adoption. Source

7. Claude Code v2.1.152 adds /reload-skills, a MessageDisplay hook, and automatic fallback model switching

Anthropic. Claude Code v2.1.152 shipped a set of additions centered on extensibility. /code-review --fix now applies findings directly to the working tree (also reachable via /simplify), and skills and slash commands can declare disallowed-tools in frontmatter to restrict tool access at runtime. A /reload-skills command plus SessionStart hook support lets users re-scan skill directories without restarting a session, and SessionStart hooks can now set session titles via hookSpecificOutput.sessionTitle. The release also introduces a MessageDisplay hook event for transforming or hiding assistant text, a pluginSuggestionMarketplaces managed setting for admin marketplace allowlisting, and automatic switching to --fallback-model when the primary model is unavailable. Source