Claude Code Hardens Enterprise Controls as Anthropic Opens Up Its Safety and Research Work

Claude Code Tightens the Controls Enterprises Asked For

Across three point releases (2.1.247, 2.1.248, 2.1.251) Claude Code added a set of features aimed squarely at organizations running it at scale. A new --restricted flag (or CLAUDE_CODE_RESTRICTED=1) strips out the tools that execute commands or code and WebFetch unless explicitly named, confines file tools to the working directory, refuses bypassPermissions, and ignores user, project, and local settings files — a locked-down mode for higher-stakes or less-trusted sessions. Alongside it came administrative visibility: a spend-limit bar in /usage and a matching rate_limits.spend_limit status-line field for developers behind a gateway with spend caps, plus a --client-label option so fleets of self-hosted runners can be identified individually instead of by hostname.

The same releases extended what sessions report and how they can be intervened on. New PreModelSwitch and PostModelSwitch hook events let scripts block, confirm, or annotate a mid-session model switch; SessionStart resume hooks now surface how stale a resumed session is and what re-caching it will cost; and Remote Control clients can watch a foreground subagent's tool calls and results stream live rather than seeing only a status indicator, as before. A SendFeedback tool introduced in 2.1.247 lets Claude draft an in-session feedback report for a user to review before sending. Read together, the changes are incremental hardening of Claude Code as an automation surface — more restriction options, more cost visibility, more auditability of what a session actually did.

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

Can AI Systems Help Police Their Own Safety Research?

Anthropic's Alignment Science Blog published two posts probing a related question from different angles: how much of alignment work itself can be delegated to AI systems. "Automated Researchers Can Reliably Mitigate Alignment Failures" addresses whether automated, AI-driven research processes can catch and correct alignment failures reliably enough to be relied on inside Anthropic's own safety pipeline. "TASTE: Can AI Models Judge AI Safety Research Proposals?" asks a narrower, adjacent question — whether AI models can meaningfully evaluate the quality of safety research proposals, a task normally reserved for human reviewers.

Both posts extend a thread Anthropic has pursued for some time: using models to help supervise or evaluate other models' work as a way to scale safety oversight alongside growing capability. The material available here states only the topic and framing of each post, not its findings or caveats, so what conclusions Anthropic draws — and with what limitations — is not established from this alone.

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

Widening the Aperture: Scientists, Hardware, and Outside Researchers

A cluster of posts published the same day points toward the same broader movement: giving outside parties more access to Anthropic's tools, infrastructure thinking, and usage data. "Expanding our support for scientists" signals additional resources for researchers using Claude in scientific work, while "Enabling independent research on how people use Claude" describes a path for external researchers to study Claude usage patterns rather than relying solely on Anthropic's internal analysis.

Two further posts extend that opening in different directions: "Previewing the Model Hardware Standard" introduces a proposed standard concerning the hardware models run on, and "Beneficial Deployments" lays out Anthropic's framing for what it considers a beneficial deployment of its models. As with the alignment posts above, the material provided states each post's subject but not its substantive content or conclusions.

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

Brief Service Disruption, Attributed to an Upstream Provider

Anthropic's status page reported, and later resolved, an incident of elevated errors affecting Claude Cowork and Claude Code on the web. By Anthropic's own account, the cause was an issue with an upstream cloud provider rather than a fault in Claude Code or Cowork themselves; the practical effect was that some sessions failed to start or disconnected mid-task, with retrying the affected session offered as the workaround. Anthropic said it remained in contact with the provider while working toward a fix, and the incident was marked resolved the same day — a short, externally-caused reliability lapse rather than a substantive product issue.

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