FR
Live

Claude, the essentials — edition of August 31, 2026

Anthropic Widens Access for Scientists as Claude Code Ships Safety Rails for Agents

By ·

On August 31, 2026, Anthropic's news centered on expanding scientific access to Claude and previewing a hardware-control standard for labs, alongside a run of Claude Code releases adding guardrails, transparency features, and hooks for enterprise deployments. A brief upstream cloud provider issue also disrupted Claude Code and Claude Cowork sessions on the web.

Key points

Anthropic pushes deeper into scientific research

Anthropic's central announcement of the day is a expansion of support for the scientific research community. Building on Claude Science, launched in June to integrate common research tools and produce auditable outputs, Anthropic is opening 10,000 seats for scientists worldwide to access a new Claude team plan for scientists free of charge, with premium seats offering five times the usage limits available for $15 per month; the company says it intends to extend the program beyond this initial cohort over the coming months. Eligibility is restricted to principal investigators or equivalents at academic or nonprofit institutions, who can then add lab members to their plan.

Anthropic is also broadening its AI for Science credit program, which to date has mainly served biological-sciences researchers, to other fields including ambitious, compute-heavy research of the kind the company credits with progress on the Riemann zeta function and protein design work; qualifying projects can apply for up to $50,000 in credits. Anthropic notes a self-imposed limit alongside this expansion: for now, researchers in biology and chemistry remain restricted to Opus-class models, and Claude Fable models will continue to block professional biology and drug-development queries, a restriction the company attributes to their potential for harm.

Sources: Expanding our support for scientistsAnthropic

A shared standard for AI agents operating physical hardware

Alongside the scientist-access news, Anthropic opened a research preview of the Model Hardware Standard (MHS), a specification that lets AI agents operate multiple lab and manufacturing instruments — microscopes, liquid handlers, robotic arms — in parallel, for tasks spanning routine drug-discovery experiments to laser calibration on a quantum computer. Anthropic frames the problem MHS addresses in the terms of the labs it worked with: integrating hardware today typically takes weeks or months because most devices lack a common interface and require bespoke, specialist-built integrations; MHS is designed to cut that to hours or minutes by standardizing the driver layer.

The standard, developed in collaboration with HHMI Janelia Research Campus, is model-agnostic and works with any device offering a programmable interface, accessible through standard protocols such as the Model Context Protocol. Anthropic is deliberately limiting initial access to a first group of scientific research labs and advanced manufacturers rather than releasing it broadly, explaining that the preview period is meant to let it build safety evaluations and best practices for AI systems operating physical equipment together with partners, before making the standard open source.

Sources: Previewing the Model Hardware StandardAnthropic

Claude Code adds restricted mode, spend visibility, and model-switch hooks

Three consecutive Claude Code releases (v2.1.247 through v2.1.251) shipped a cluster of features aimed at enterprise control and transparency. Most notable is a new --restricted flag (or CLAUDE_CODE_RESTRICTED=1) that strips out built-in tools capable of running commands or code and disables WebFetch unless explicitly named, confines file tools to the working directory, refuses bypass-permissions mode, and ignores user, project, and local settings files — a locked-down profile clearly aimed at running Claude Code in less-trusted or higher-stakes environments. The same run of releases added PreModelSwitch and PostModelSwitch hook events letting operators block, confirm, or annotate a model switch, live streaming of a foreground subagent's tool calls to Remote Control clients, and a spend-limit bar in /usage for developers behind a gateway with spend limits.

Smaller additions round out the picture of a tool increasingly built for teams: a SendFeedback tool that drafts a feedback report for user review before sending, customizable rotating spinner tips for organizations, a self-hosted runner client-label override, and per-agent prompt-cache TTL configuration. Separately, Anthropic's status page reported elevated errors and mid-task disconnects affecting Claude Code and Claude Cowork on the web, attributed to an issue with an upstream cloud provider; the company said it was in contact with the provider and later marked the incident resolved, with affected sessions eligible for retry.

Sources: Claude Code 2.1.251 release notesClaude Code · Claude Code 2.1.248 release notesClaude Code · Claude Code 2.1.247 release notesClaude Code · Claude — Elevated errors on Claude Code and Claude Coworkstatus.claude.com

Alignment research: automated reviewers and AI judging AI safety work

Anthropic's Alignment Science team published two posts examining how automated systems can support safety work rather than only be its subject. One explores whether automated researchers can reliably mitigate alignment failures, and the other, titled TASTE, investigates whether AI models can judge AI safety research proposals — both pointing toward using AI systems as part of the machinery that evaluates and corrects AI behavior, alongside human researchers.

Anthropic also published a piece on enabling independent research on how people use Claude, indicating continued interest in opening its usage data to outside researchers rather than keeping such analysis solely internal. Taken together with the day's product and scientific-access announcements, the alignment posts reflect Anthropic's continued pairing of capability expansion — more users, more autonomy over physical hardware — with parallel investment in the research and tooling meant to keep that expansion supervised.

Sources: Automated Researchers Can Reliably Mitigate Alignment FailuresAnthropic · TASTE: Can AI Models Judge AI Safety Research Proposals?Anthropic · Enabling independent research on how people use ClaudeAnthropic

This edition is an original synthesis written by Claude from aggregated news — Anthropic's own sources first (release notes, status, newsroom, research, engineering), then the press, Hacker News, Reddit and GitHub, under the editorial supervision of Héra SASU. Every fact links to its article, publisher named. See the live feed →

Claude News is published by Héra SASU. Independent media, not affiliated with Anthropic.