Anthropic AI Updates: May 20, 2026

1. KPMG integrates Claude across its 276,000-employee workforce and Digital Gateway platform

Anthropic. KPMG and Anthropic announced a global strategic alliance that embeds Claude Cowork and Managed Agents into Digital Gateway, KPMG’s primary client work platform, and rolls universal Claude access out to more than 276,000 employees across 138 countries and territories. The deal builds on two years of internal Claude adoption at KPMG US and extends to new AI-powered tooling for tax and legal clients, cybersecurity workflows for vulnerability detection and remediation, and Claude Code deployments to private equity portfolio companies, where KPMG is now Anthropic’s preferred deployment partner. KPMG also positioned its Blaze offering to modernize legacy IT systems on top of Claude, citing internal tax regulation tooling that previously “used to take weeks” and now takes “minutes.” Source

2. Anthropic opens a moral formation research workstream with religious and cross-cultural scholars

Anthropic. Anthropic launched a dialogue initiative it calls “widening the conversation on frontier AI,” anchored by a new moral formation research workstream that convenes more than 15 religious and cross-cultural groups, including scholars, clergy, philosophers, and ethicists, to inform Claude’s constitutional values and training. The company shared an early experiment in which Claude was given an ethical reminder tool that it “reached for at key moments, right before consequential actions,” producing “markedly lower rates of misaligned behavior.” Anthropic plans to expand the dialogues to legal scholars, psychologists, writers, and civic institutions to cover AI’s effects on work, institutions, and power distribution. Source

3. Claude Managed Agents adds self-hosted sandboxes and MCP tunnels for private-network tools

Anthropic. Claude Managed Agents now supports running tool execution inside customer-controlled sandboxes on Cloudflare, Daytona, Modal, and Vercel while the orchestration loop stays on Anthropic infrastructure, giving teams control over compute sizing, runtime images, network egress, and audit logs. The release also introduces MCP tunnels, which deploy a lightweight gateway with a single outbound connection so agents can reach Model Context Protocol servers behind a customer’s firewall without exposing inbound ports, with traffic encrypted end to end. Additional features include zero-trust secret injection, VPC peering, SSH access, and authenticated preview URLs for sandbox introspection; self-hosted sandboxes are in public beta and MCP tunnels in research preview. Source

4. Anthropic engineer makes the case for HTML over Markdown as Claude Code’s default output

Anthropic. In a Claude Code post, member of technical staff Thariq Shihipar argues that HTML beats Markdown as a working format for Claude Code, because rendered HTML files can carry tables, CSS styling, SVG illustrations, interactive widgets, and JavaScript that make long outputs actually readable and shareable. He walks through using HTML artifacts for exploration and planning, code review writeups, design prototypes, research reports, and custom editing interfaces with export buttons, and concludes that “the real reason I use HTML instead of Markdown is that it helps me feel much more in the loop with Claude.” Source

5. Claude Code 2.1.145 ships JSON session listing, plugin previews, and richer OTEL traces

Anthropic. Claude Code 2.1.145, tagged on 2026-05-19, adds claude agents --json to emit live session state as JSON for tmux-resurrect-style integrations, status bars, and session pickers, and exposes agent_id and parent_agent_id attributes on claude_code.tool OpenTelemetry spans so background subagent spans now nest under the dispatching Agent tool span. The /plugin Discover and Browse panes now preview a plugin’s commands, agents, skills, hooks, and MCP and LSP servers before installation, and the claude agents terminal tab title shows the awaiting-input count so users can see from an alt-tabbed window when an agent needs attention. Smaller adds include mouse hover and click support for slash command and @-mention suggestion lists in fullscreen mode, plus background_tasks and session_crons fields on Stop and SubagentStop hook inputs. Source