Anthropic AI Updates: August 7, 2026

1. Anthropic Refines Fable 5 Biology Safeguards to Cut False Positives

Anthropic. Anthropic detailed updates to the safety classifiers that govern Fable 5, its frontier model capable of outperforming experts on complex biological tasks, saying it retrained the system to roughly 85% fewer biology-related fallbacks. The change means everyday health and educational questions, such as interpreting lab results or understanding symptoms, are less likely to be redirected to the less-capable Opus 5 model, while Fable 5 continues to block dual-use requests in areas like virology, toxicology, and molecular design outside trusted pathways. Anthropic said the retrained classifier drew on feedback from external experts to better separate benign queries from genuinely dangerous research. Source

2. Millennium and Anthropic Build a Digital Risk Analyst with Claude

Anthropic. Anthropic and hedge fund Millennium announced they are co-developing a “digital risk analyst,” an AI system built on Claude and Claude Code that helps risk managers analyze positions across asset classes while keeping human judgment central. The tool logs its reasoning, tests actions in sandboxed environments, and requires expert approval before any implementation, and it can retain context over time to explain daily risk changes. Anthropic said Millennium’s 340-plus investment teams already use Claude Code, and its risk experts are building the analyst alongside Anthropic’s research and applied AI teams in Millennium’s AI lab. Source

3. Claude Code Adds Self-Hosted Environments in Public Beta

Anthropic. Anthropic launched a public beta of self-hosted environments that let organizations run Claude Code sessions on their own infrastructure, so sessions execute inside a company’s network and can reach internal services, databases, and registries without exposing them to the public internet. Repository checkouts, build artifacts, and created files stay on the organization’s infrastructure while conversation and inference continue on Anthropic’s servers, and teams can choose fixed or on-demand runner modes to manage capacity. The feature is available to Claude Team and Enterprise plan organizations, excluding those on zero-data-retention agreements, and requires dedicated engineering resources to set up and operate. Source

4. Claude Code v2.1.223 Adds Marketplace Wildcards and Model Restriction Warnings

Anthropic. Claude Code v2.1.223 introduces support for owner/* wildcard entries in the strictKnownMarketplaces and blockedMarketplaces managed settings, letting administrators allow or block every marketplace repository under a GitHub organization at once. The release also adds a warning when workflow agents, forked skills, slash commands, or resumed background agents request a restricted model and must fall back to the parent model, plus a hint on continuing a cloud session locally via claude --teleport <session id>. Source