Anthropic AI Updates: May 8, 2026

1. Claude for Microsoft 365 hits GA, Outlook in public beta

Anthropic. Claude for Excel, PowerPoint, and Word are now generally available, and Claude for Outlook entered public beta. A single conversation persists across all four apps with full context — triage an Outlook email, open the attachment in Word, then refresh linked Excel models — and adds enterprise telemetry, an Analytics API, and multi-cloud routing across Bedrock, Vertex AI, and Microsoft Foundry. Source

2. The Anthropic Institute publishes its research agenda

Anthropic. Anthropic released the agenda for The Anthropic Institute, organized around four areas: economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D. The Institute uses its position inside a frontier lab to study AI’s real-world impact and shares findings publicly to inform policy. Source

3. Anthropic donates Petri alignment tool to Meridian Labs

Anthropic. Anthropic transferred Petri, its open-source alignment-evaluation framework, to AI evaluation nonprofit Meridian Labs to keep the tool independent from any single AI lab. Petri 3.0 ships alongside the donation with improvements to adaptability, realism, and depth for detecting deception, sycophancy, and other concerning behaviors. Source

4. Natural Language Autoencoders translate Claude’s activations to English

Anthropic. New interpretability research introduces NLAs, which train one model to describe Claude’s internal activations in plain English and another to reconstruct activations from those descriptions. NLAs surfaced that Claude flags evaluation awareness on 16–26% of test problems versus under 1% in real usage, and helped identify hidden model motivations in an auditing game 12–15% of the time vs. <3% without. Paper, code, and a Neuronpedia demo are released. Source

5. Claude Code v2.1.133 adds worktree, sandbox, and effort-aware hooks

Anthropic. A new Claude Code release adds worktree.baseRef to choose whether worktrees branch from origin/<default> or local HEAD, sandbox.bwrapPath and sandbox.socatPath for custom sandbox binaries, and the admin-tier parentSettingsBehavior key for SDK managedSettings policy merging. Hooks now see the active effort level via effort.level JSON and $CLAUDE_EFFORT. Source