Claude Expands Into Government and Enterprise as Anthropic Faces New Trust Questions

Inside the model: a hidden reasoning space, and its blind spots

Anthropic disclosed that its researchers identified an internal reasoning zone within Claude that appears to have emerged on its own rather than having been deliberately engineered — a finding that matters less as a product feature than as a data point for the company's interpretability work, its long-running effort to understand what actually happens inside the models it ships. That same research impulse extends outward: Anthropic is also exploring whether Claude can help accelerate the search for new medicines, a use case that depends on the model's reasoning holding up under scientific scrutiny.

The limits of that reasoning were also on display today. A fundamental security flaw known as "Bad Epoll" reportedly went undetected by Anthropic's AI, a reminder that better introspection into how a model reasons does not automatically translate into catching every real-world failure — a caveat worth holding alongside the CISA news below.

Sources: Anthropic découvre un espace de pensée caché au sein de Claude · Anthropic veut que Claude aide à inventer les médicaments de demain · Une vulnérabilité « Bad Epoll » échappe à l'analyse de l'IA d'Anthropic

Claude moves off the desktop and gets a memory of its own

Claude Cowork, previously confined to a desktop application, is now reaching web browsers and mobile devices, part of a broader pattern of Anthropic loosening the constraints on where and how people access Claude. It arrives alongside a separate but complementary change: Claude Chat's new memory feature, which organizes what the assistant knows about a user into distinct, editable notes rather than an opaque running log — giving users visibility and control over what persists across conversations.

Two smaller product signals point the same direction. Anthropic introduced a Learning Mode designed to guide users toward understanding rather than simply supplying answers, and it quietly extended expanded access to Fable 5 for paying subscribers through July 12 without explanation, a delay that has drawn notice in the user community. Claude Code, meanwhile, received routine patches (versions 2.1.203 and 2.1.204) fixing hook and session-authentication bugs — unglamorous but a sign of steady iteration on the developer tooling.

Sources: Claude Cowork disponible sur mobile et web, au-delà du bureau · Claude Chat introduit des fichiers de mémoire organisés et modifiables · Claude lance un « mode apprentissage » pédagogique · L'accès étendu à Fable 5 prolongé jusqu'au 12 juillet · Claude Code : correctifs dans les versions 2.1.203 et 2.1.204

Demand meets doubt: Anthropic's standing with governments and enterprises

Anthropic's institutional footprint kept growing today even as one of its biggest backers appears to be hedging. Bloomberg reports Microsoft is reducing its reliance on Anthropic and OpenAI in favor of models developed in-house, a shift that would matter given Microsoft's role as both investor and distribution partner. At the same time, the US government is going the other direction: despite budget cuts, the CISA and the Trump administration are using Anthropic's Mythos model to audit government code for security flaws, and IT services group Globant has struck a partnership to help enterprises deploy Claude more broadly.

This mixed picture matches an analysis from L'Usine Digitale describing Anthropic as a vendor enterprises are eager to adopt yet remain wary of trusting fully — a tension that now extends into politics, with an Anthropic-linked controversy reportedly becoming a talking point in presidential campaign debates over AI's role in society. Taken together, the day's business news suggests Anthropic's commercial and government reach is expanding even as scrutiny of the company keeps pace.

Sources: Microsoft remplacerait Anthropic et OpenAI par ses propres modèles · Le gouvernement américain utilise Claude Mythos pour auditer la cybersécurité de ses sites · Globant s'associe à Anthropic pour déployer Claude en entreprise · Anthropic, un fournisseur plébiscité mais qui inspire la méfiance

From the community: skepticism and a small security win

Reaction to the Cowork rollout has not been uniformly positive: some users are openly questioning what Cowork actually adds over the existing Claude Code Desktop application, a sign that Anthropic will need to articulate a clearer distinction between the two products as both continue to evolve. On a smaller but telling note, one user reported that Fable 5 flagged a suspicious PowerShell persistence entry on their PC during a routine check — an informal, single anecdote, but one that illustrates the kind of everyday utility Anthropic is betting on as it pushes Claude into more corners of daily computing.

Sources: Cowork face à Claude Code : la communauté s'interroge sur son utilité · Un utilisateur affirme que Fable 5 a détecté un malware sur son PC