Anthropic AI Updates: July 18, 2026

1. Cursor Reports Claude Fable 5 Sets a New High on Its Internal Benchmark

Anthropic. Anthropic published a case study describing how Cursor evaluated Claude Fable 5 on CursorBench, an internal benchmark built from ambiguous, underspecified prompts meant to mirror real developer work, where the model scored 72.9 percent at maximum effort and set a new high. Cursor highlighted the model’s global reasoning, citing a space-flight simulator test in which Fable 5 completed a moon landing in hours after running its own orbital test flight first, compared with Claude Opus taking more than 12 hours without success. A Cursor engineer said the model held onto goals and context without needing repeated reminders across long tasks. Source

2. Anthropic Deputy CISO Publishes a Risk Framework for Agentic AI

Anthropic. Anthropic Deputy CISO Jason Clinton published guidance arguing that security leaders should abandon the goal of zero risk and instead make agentic AI risks legible and bounded through deliberate risk acceptance. The piece offers a four-question assessment covering untrusted content ingestion, an agent’s actions and identity, blast radius, and observability, along with seven controls such as identity from an existing identity provider, connector allowlists, per-action approvals, sandboxed execution, egress allowlisting against prompt injection, telemetry streamed to a SIEM via OpenTelemetry, and organization-wide kill switches. Its core principle is to grant the narrowest capability that still completes the task. Source

3. Claude Code Adds Background Forks, Runaway-Loop Caps, and an End-Conversation Tool

Anthropic. Anthropic shipped Claude Code v2.1.212 and v2.1.214, adding several new features. v2.1.212 changes /fork to copy a conversation into a new background session while the former in-session subagent becomes /subtask, and it adds session-wide caps on WebSearch calls and subagent spawns, each defaulting to 200, to stop runaway loops, a claude auto-mode reset command, automatic backgrounding of MCP tool calls running longer than two minutes, and a /resume session picker. v2.1.214 adds an EndConversation tool that lets Claude end sessions with highly abusive users or jailbreak attempts, a progress heartbeat for long-running tool calls, and permission prompts for docker commands carrying daemon-redirect flags. Source

4. Epilepsy Foundation Builds a Claude-Powered Assistant Grounded in Reviewed Content

Anthropic. Anthropic published a case study on the Epilepsy Foundation, which built Sage, a conversational assistant on epilepsy.com that grounds its answers in a knowledge base of 3,200 medically reviewed pages and runs on Claude through Amazon Bedrock with HIPAA compliance. The Foundation selected Claude after testing five AI systems on accuracy, empathy, tone, and recommendations, and it can deploy new model versions in about 30 minutes. Sage logged roughly 60,000 interactions in its first year with minimal promotion and raised interactions per session from 1.2 to about 3. Source

5. Eve Legal Says Claude Helps Plaintiff Firms Settle Cases Up to 60 Days Faster

Anthropic. Anthropic detailed how Eve Legal, an AI platform for plaintiff-side law firms, uses Claude to analyze case files of up to 70,000 pages, draft demand letters in each firm’s style, build medical chronologies from unstructured records, and run agents that monitor thousands of cases nightly. The company reported that cases settle up to 60 days faster, settlement values rise up to 30 percent, and lawyers handle up to 30 additional cases per year, while the platform processes about 12.5 million documents monthly across more than 1,700 firms. Eve said it automated 80 percent of previously manual work within the first 90 days of a deployment. Source

6. National Domestic Workers Alliance Deploys a Bilingual Claude Advisor

Anthropic. Anthropic published a case study on the National Domestic Workers Alliance, which built Ask Aya, a Claude-based tool that gives domestic workers 24/7 guidance in English and Spanish on issues like pay negotiation and workplace disputes. The tool routes queries across Claude Opus, Sonnet, and Haiku under a zero-data-retention agreement and refers sensitive legal and immigration questions to human experts. In its beta, 93 percent of testers applied the advice, 25 percent negotiated pay increases, and 61 percent reported greater influence over their compensation. Source