Anthropic Widens Claude's Reach — Into Labs, Safety Research, and Tighter Code Controls

Claude Code tightens sandboxing and cost visibility

Across three point releases issued in close succession, Anthropic added a --restricted mode (also settable via CLAUDE_CODE_RESTRICTED=1) that strips out tools capable of running commands or code and blocks WebFetch unless explicitly named, confines file tools to the working directory, refuses bypassPermissions, and ignores user, project and local settings files — effectively a lockdown mode for running Claude Code in lower-trust contexts. The same set of updates added a SendFeedback tool that drafts an in-session feedback report for the user to review and send from /feedback, and a tip surfaced on Bash permission prompts pointing users toward auto mode with a one-keystroke opt-in.

The releases also expanded operational visibility for teams running Claude Code at scale: /usage gained a spend-limit bar and a rate_limits.spend_limit status-line field for developers behind a Claude apps gateway with spend caps, agent frontmatter can now set a per-agent prompt-cache TTL, and SessionStart resume hooks now report session staleness and the estimated cost of re-caching. New PreModelSwitch/PostModelSwitch hook events let integrators block, confirm, or annotate automatic model switches, and Remote Control clients can now watch a foreground subagent's tool calls and results stream live rather than seeing status only (background subagents remain status-only). Together the changes read as a coordinated push toward stricter sandboxing and clearer cost accounting rather than new user-facing features.

Sources: Claude Code 2.1.251 release notes — Claude Code · Claude Code 2.1.248 release notes — Claude Code · Claude Code 2.1.247 release notes — Claude Code

A short outage, resolved

Anthropic's status page reported and then resolved an incident of elevated errors affecting Claude Cowork and Claude Code on the web. Anthropic attributed the disruption to an issue with an upstream cloud provider rather than to Claude Code itself: some sessions failed to start or disconnected mid-task, affected sessions could be retried, and Anthropic said it remained in contact with the provider while working toward a fix. The episode underscores that Claude's web-based coding surfaces depend on third-party infrastructure outside Anthropic's direct control.

Sources: Elevated errors on Claude Code and Claude Cowork — status.claude.com

Alignment Science asks whether automated judgment can be trusted

Anthropic's Alignment Science blog published two posts probing the reliability of AI-assisted safety work. "Automated Researchers Can Reliably Mitigate Alignment Failures" addresses whether automated, AI-driven researchers can dependably catch and fix alignment problems, while a companion piece, "TASTE: Can AI Models Judge AI Safety Research Proposals?", asks whether models can be trusted to evaluate the merit of safety research itself. Both questions bear on the same underlying premise in Anthropic's approach to scaling alignment work with AI collaborators: that effort only holds up if the automated judgments involved are themselves sound, which is precisely what the two posts set out to examine.

Sources: Automated Researchers Can Reliably Mitigate Alignment Failures — Anthropic · TASTE: Can AI Models Judge AI Safety Research Proposals? — Anthropic

From chat to lab bench: the Model Hardware Standard

Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification that lets AI agents operate physical lab and manufacturing equipment — microscopes, liquid handlers, robotic arms — in parallel, developed initially with HHMI Janelia Research Campus. Anthropic frames the underlying problem as fragmented integration: most lab devices don't natively communicate with each other, and wiring them together today typically takes specialists weeks or months of bespoke work; MHS standardizes the driver layer so that work can shrink to hours, works with any device exposing a programmable interface, and is model-agnostic, reachable by any agent harness through standard protocols such as MCP. The preview is limited to a first group of scientific research labs and advanced manufacturers, and Anthropic said the early release to partners across science, robotics, electronics and manufacturing is meant to build safety evaluations and best practices before the standard is eventually open-sourced.

The hardware standard arrived alongside other posts grouped under Anthropic's "Beneficial Deployments" heading, including pieces on expanding support for scientists and on enabling independent research into how people actually use Claude. Read together, the throughline is Anthropic extending Claude's reach beyond chat and code — into physical experimentation and into external researcher access to usage data — while pairing each expansion with an explicit safety-and-evaluation framing rather than presenting it as a straightforward capability launch.

Sources: Previewing the Model Hardware Standard — Anthropic · Expanding our support for scientists — Anthropic · Enabling independent research on how people use Claude — Anthropic · Beneficial Deployments — Anthropic