The AI week, distilled.
Week 28 · 2026
This week in Anthropic

Anthropic extends Claude Cowork access while publishing new safety interpretability research

Anthropic expanded Claude Cowork availability to web and mobile and extended higher usage limits into early August. Separately, it published interpretability research describing an internal mechanism intended to help control dual-use knowledge.

01

Claude Cowork rolls out to web and mobile

Anthropic said Claude Cowork is rolling out on web and mobile, with sessions and files syncing across devices. Anthropic also said it is extending doubled Cowork usage limits through August 5.

  • Cross-device continuity supports executives and distributed teams that move between laptop and phone during the workday.
  • The temporary higher limits affect pilot design, because usage ceilings can become the main constraint in agent-style workflows and large-file tasks.
  • Mobile and web availability reduces friction for field and frontline scenarios where desktop-only assistants are not practical.
02

Anthropic details interpretability “off switch” research

Anthropic published research describing an internal “mental workspace” in Claude that can hold thoughts not shown in the output. Anthropic framed the work as relevant to controlling dual-use knowledge and improving safety controls.

  • Safety and interpretability work can shorten vendor-risk assessments for regulated Czech buyers who need evidence of control mechanisms beyond policy statements.
  • If productized, this line of research could translate into stronger governance features such as safer tool-use constraints and better audit narratives for internal compliance.
  • Procurement teams can use the paper to benchmark Anthropic’s safety posture against competitors when selecting a model provider for high-impact use cases.