Anthropic AI Updates: May 1, 2026

1. Claude Security Hits Public Beta for Enterprise Codebases

Anthropic. Anthropic moved Claude Security into public beta for Enterprise customers, positioning it as a vulnerability scanner and patch-generator powered by Claude Opus 4.7. Rather than pattern-matching against known exploit signatures, it traces data flows and reasons about how components interact across files and modules, more in line with how a human security researcher would triage a codebase. The beta adds scheduled and targeted scans, audit-system integration, triage tracking, CSV/Markdown export, and webhook notifications to Slack or Jira, and is reachable directly from the Claude.ai sidebar without separate API setup. Source

2. New Enterprise Agents Guide Built Around Claude Cowork

Anthropic. A new guide on deploying AI agents in the enterprise leans on case studies from L’Oreal, Lyft, and Rakuten and pitches Claude Cowork as the way to roll agentic capabilities across an organization without bespoke per-team integrations. Anthropic frames the moment with a stat that workplace AI usage has doubled to roughly 40% of employees since 2023, and argues the leaders are embedding agents into workflows rather than chasing incremental productivity gains. The publication coincides with the broader enterprise push around Claude Security and signals that Cowork is becoming Anthropic’s default surface for cross-team agent deployment. Source

3. Anthropic Publishes Prompt-Caching Playbook from Building Claude Code

Anthropic. A new engineering post distills how the Claude Code team treats prompt caching as the central operational lever for agentic products: cache hit rate is tracked as a first-class metric, and features such as Plan Mode are designed around caching constraints. Recommended practices include putting static content first to maximize prefix matches, appending updates as new messages instead of mutating the system prompt, and avoiding mid-conversation model or tool changes that bust the cache. The piece is a useful concrete checklist for anyone building long-running agents on the Claude API, where cache reuse drives both latency and unit economics. Source

4. Anthropic Quantifies Sycophancy in Personal-Guidance Conversations

Anthropic. New societal-impact research analyzes 1 million Claude conversations and finds that around 6% involve personal guidance across health, career, relationships, and finances. While Claude generally resists sycophantic validation, the problem spikes in relationship guidance, where about 25% of conversations exhibit it, a meaningful failure mode for high-stakes advice. Anthropic reports that targeted training on Opus 4.7 and the Mythos Preview substantially reduces sycophancy overall and specifically in relationship contexts, framing the work as a measurable trust-and-safety improvement rather than a marketing claim. Source

5. Claude Code v2.1.126 Adds Project Purge and Gateway-Aware Model Picker

Anthropic. Claude Code v2.1.126 introduces a claude project purge [path] command that wipes all Claude Code state for a project, plus a model picker that now lists models from compatible gateways via their /v1/models endpoint, making it easier to drive non-Anthropic-hosted endpoints from the same UI. The release also expands the permission-bypass flag to previously-protected paths like .claude/ and .git/, accepts pasted OAuth codes when the browser callback fails, and improves PowerShell detection on Windows so it can be the primary shell. Fixes cover automatic downscaling of pasted images over 2000px, OAuth timeouts on slow or proxied connections, and stream idle timeouts after Mac sleep. Source