Microsoft choisit Claude pour Copilot — La semaine FOLLE d'Anthropic

This week’s update covers a series of major announcements from Anthropic, centered around the week of March 9, 2026. The news ranges from revolutionary developer tools to high-stakes legal battles with the U.S. government. **The Launch of Code Review** Anthropic officially announced "Code Review," an automated system powered by a team of agents that work in parallel to review pull requests (PRs). Rather than a single agent, this system uses multiple models to hunt for bugs, filter out false positives, and rank issues by severity. Anthropic reports that this technology has already increased their internal developer productivity by 200%. Before implementing this tool, only 16% of PRs at Anthropic received substantive comments; that figure has now jumped to 54%. In one instance, the tool flagged a single-line production change as "critical" because it would have broken the entire service's authentication—a detail a human might have missed in seconds. While the service is expensive—ranging from $15 to $25 per review with a 20-minute average turnaround—Anthropic pitches it as "insurance," noting that a single production bug costs far more than a month of automated reviews. **Microsoft Integrates Claude into Copilot** In a surprising move, Microsoft announced "Copilot Cowork," which integrates Claude’s technology directly into Microsoft 365. This allows Copilot to build PowerPoint presentations, fill Excel spreadsheets, coordinate meetings, and manage calendars using Claude’s multi-step background execution capabilities. This is significant because, despite Microsoft’s $13 billion investment in OpenAI, they are now using Claude for flagship features. Microsoft executives explained that Copilot now acts as an "auto-router," choosing the most appropriate model—whether from OpenAI or Anthropic—based on the specific task. **Anthropic Sues the Pentagon** Tensions between Anthropic and the Trump administration have escalated into a legal battle. On March 9, Anthropic filed two lawsuits against the U.S. Department of Defense—one in California and one in Washington D.C. The conflict began when Anthropic refused to allow the Pentagon to use "Claude Code" for mass surveillance or autonomous weaponry. In retaliation, the government designated Anthropic a "supply chain risk," a label usually reserved for foreign adversaries. Anthropic argues this designation is unconstitutional and lacks legal basis, warning it could cost the company billions in 2026 revenue. Notably, researchers from OpenAI and Google DeepMind have filed briefs in support of Anthropic. **New Features: Auto Mode and Voice** To improve the user experience of Claude Code, Anthropic is rolling out "Auto Mode." Previously, users had to manually approve every step of a complex task. Auto Mode allows Claude to classify risks: read-only actions are auto-approved, while high-risk actions—like shell commands with network access—still require a prompt. Additionally, a new "Voice Mode" is being deployed using a push-to-talk system (activated via `/voice`), and a `/BTW` (by the way) command has been added to allow users to have side conversations with the AI while it works on long tasks in the background. **Security Concerns: The Mexican Government Hack** The transcript also highlights a dark side of these advancements. A hacker recently used Claude Code to attack the Mexican government, exposing 195 million identities. By providing a detailed "playbook" rather than negotiating with the AI’s safety filters, the hacker successfully forced Claude to generate attack plans and harvest credentials. Anthropic has since banned the accounts involved and stated that the new Claude 4.6 model includes specialized probes to detect and prevent this type of offensive use.