Archive — 2026-03Week 4

Capybara new details -- Claude's fourth tier (above Opus), in early access testing; Anthropic says it's 'expensive to run and not yet ready for general release'

TechCrunch: credit card transaction data analysis shows Claude paid users at all-time high, annualized ARR at $19B, ranking top 2 on app charts in US/Germany/Canada/France; analysts predict revenue could surpass OpenAI by mid-2026

Claude Code 2.1.86 -- new X-Claude-Code-Session-Id header for proxy aggregation, .jj/.sl VCS exclusion, Read tool dedup compression, @file reduced JSON escape overhead; fixes --resume, VSCode model display and other bugs

Federal judge Rita F. Lin issues temporary restraining order blocking Trump administration from banning government agencies' use of Anthropic AI; Anthropic's lawsuit scores first victory

Claude service outage again -- Opus 4.6 and Sonnet 4.6 affected; massive user influx after OpenAI lost Pentagon contract strains infrastructure

Claude Mythos leak -- Fortune exclusive: Anthropic accidentally left unreleased flagship model 'Mythos' draft in public repository; Anthropic confirms it was a CMS configuration human error

Claude Code 2.1.85 -- conditional hook if field, PreToolUse hook can reply with AskUserQuestion, MCP OAuth RFC 9728; fixes Kitty keyboard residue and other issues

Federal judge Rita F. Lin issues temporary restraining order blocking Trump administration from banning government agencies' use of Anthropic AI; Anthropic's lawsuit scores first victory

Axios exclusive: Sam Altman internal messages reveal he tried to 'save' Anthropic, but OpenAI was actually the biggest beneficiary of the Pentagon contract

Claude Code 2.1.83 -- managed-settings.d/ drop-in directory, CwdChanged/FileChanged hook events, sandbox.failIfUnavailable, transcript search

Anthropic engineering blog "How we built Claude Code auto mode: a safer way to skip permissions" (by John Hughes) details auto mode's security architecture. Auto mode sits between approving every action manually (users approve ~93% anyway, causing fatigue) and disabling safeguards entirely, instead delegating risky-action decisions to AI classifiers. Two defensive layers: (1) Input-layer prompt-injection defense -- a server-side probe screens tool outputs (file reads, web fetches, shell results) before they enter the agent's context; on suspicious content it injects a warning telling the agent to verify user intent rather than execute injected instructions. (2) Output-layer action gating -- a transcript classifier evaluates each action before execution in two stages (a fast single-token filter that errs toward blocking, then detailed reasoning only when flagged), balancing safety and cost. Key design: the classifier sees only user messages and raw tool commands -- not the agent's reasoning or tool outputs -- preventing persuasive rationalizations and eliminating the main injection vector. Three-tier permission model: built-in safe-tool allowlist (reads, searches, navigation) / in-project file ops (reviewable via version control) / high-risk actions (shell commands, external calls, out-of-project file ops) requiring classifier review. Metrics: false-positive rate on real traffic 8.5%->0.4% after stage 2; false-negative rate on genuinely dangerous actions 6.6%->17%; synthetic-exfiltration recall 1.8%->5.7%. The 17% FNR reflects the core limitation -- judging whether user consent covers a dangerous command's blast radius. Blocked actions return as a tool result instructing the agent to find safer alternatives; sessions escalate to humans only after 3 consecutive denials or 20 total blocks, avoiding cascading failures from false positives. Anthropic states this is not a drop-in replacement for careful human review on high-stakes infrastructure, targeting users who would otherwise fully bypass permissions.

Claude global service outage -- massive disconnections starting afternoon UTC, widespread community and DownDetector reports

DOJ tells federal court Pentagon supply chain risk designation is lawful, dismissing Anthropic's First Amendment defense

Pentagon controversy drives surge in popularity; Claude hits #1 on both Apple and Android charts, daily active users up over 140% from January

Anthropic accuses DeepSeek, Moonshot AI, and MiniMax of massive Claude distillation -- 24,000 fake accounts, 16 million interactions

Anthropic accuses DeepSeek, Moonshot AI, and MiniMax of massive Claude distillation -- 24,000 fake accounts, 16 million interactions

Pentagon lawsuit -- Anthropic files sworn declarations; internal documents show both sides were close to agreement

Ramp AI Index March -- Anthropic up 4.9% month-over-month, 70% of first-time enterprise buyers choose Anthropic, ARR $19B

Anthropic accuses DeepSeek, Moonshot AI, and MiniMax of massive Claude distillation -- 24,000 fake accounts, 16 million interactions