Archive — 2026-05Week 1

Anthropic blog "Collaborate with Claude across Excel, PowerPoint, Word and Outlook" — Claude for Excel, PowerPoint, and Word Microsoft 365 add-ins are now generally available (GA), with Claude for Outlook entering public beta for all paid plans. The headline feature is automatic context carryover across apps: Claude can read from one app and make changes in another without manual copy-paste. The Outlook integration triages the inbox into "needs your response", "can be drafted", and "noise" buckets and can draft replies directly. All four integrations install via the Microsoft Marketplace and support Windows, Mac, and the web versions of Office, requiring a paid Claude plan (Pro, Team, or Enterprise). Mobile Office apps are not currently supported.

Anthropic donates its open-source alignment testing tool Petri to AI evaluation nonprofit Meridian Labs — Petri is a toolbox for testing large language models for concerning behaviors including deception, sycophancy, and cooperation with harmful requests, originally built through Anthropic's Fellows program and launched in October 2025; it has been integrated into Anthropic's Claude alignment assessments since Claude Sonnet 4.5 and adopted by external organizations such as the UK's AI Security Institute; Petri 3.0 improvements include architectural flexibility, increased realism via an add-on called "Dish" that injects real system prompts, and deeper integration with the Bloom alignment tool; Anthropic says transferring development to an independent nonprofit ensures neutral, credible results and removes association with any single AI lab — Petri joins Inspect and Scout in an open technology stack available to researchers, labs, and governments.

Anthropic publishes the four focus areas of The Anthropic Institute (TAI) — TAI's mission is to "investigate AI's impact on the world using insights from operating a frontier AI lab, then share learnings with the public," acting as a research-to-policy bridge, publishing findings, tools, and datasets, and feeding into Anthropic's own deployment decisions and Long-Term Benefit Trust framework; the four focus areas are: (1) Economic Diffusion — how powerful AI systems reshape labor markets, job creation, productivity, and wealth distribution across regions and industries; (2) Threats and Resilience — dual-use AI risks (cybersecurity, biotech) and societal safeguards against emerging threats; (3) AI Systems in the Wild — how people and institutions interact with deployed AI, including effects on epistemology, critical thinking, and human agency; (4) AI-Driven R&D — how AI assists scientific research across fields and the implications of AI systems autonomously improving themselves.

Bloomberg reports Anthropic is working to make its Claude chatbot more appealing to consumers — since late last year, Anthropic has tasked employees with improving how the chatbot handles personal queries like health questions; although originally aimed at enterprise users, recent traction on the consumer side is prompting Anthropic to invest in non-work usage scenarios; the article runs alongside the Code w/ Claude conference (5/6) disclosure of 80x year-over-year growth in Q1 2026 revenue and usage, signaling Anthropic's parallel push on both fronts.

Elevated errors on Claude Opus 4.7 — status.claude.com began investigating at 05-07 12:10 UTC, identified the issue at 12:19 UTC, entered monitoring at 12:20 UTC, and resolved at 12:45 UTC; duration approximately 35 minutes. Affected: claude.ai, Claude Console, Claude API, Claude Code, Claude Cowork.

Connection failures for organizations restricting GitHub access by IP address — status.claude.com identified the issue at 05-06 22:32 UTC: a recent infrastructure change altered the IP addresses Anthropic uses for outbound connections to GitHub, potentially affecting Claude Code remote sessions, GitHub Enterprise plugin syncing, and Claude Security functionality; Anthropic stated they are reverting the change to restore the previous IP ranges, with no allowlist changes needed on the user side; status.claude.com indicated they will post an update once the fix is deployed.

Claude Code 2.1.132 — Bash subprocess environment now includes `CLAUDE_CODE_SESSION_ID` matching the `session_id` passed to hooks; added `CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1` env var to opt out of the fullscreen alternate-screen renderer and keep the conversation in the terminal's native scrollback; added a 'Pasting…' footer hint while a Ctrl+V image paste is being read from the clipboard. Fixes: external SIGINT (IDE stop button, `kill -INT`) now runs graceful shutdown — terminal modes are restored and the `--resume` hint is printed; uncaught exception when the terminal is closed or SSH disconnects mid-session under the native build; `--resume` failing with `no low surrogate in string` after a tool error truncated an emoji (pre-corrupted sessions are sanitized on load); `--permission-mode` flag being ignored when resuming a plan-mode session with `-p --continue`/`--resume`, and plan mode not being re-applied after `ExitPlanMode` within the same session; fullscreen mode showing a blank screen after laptop sleep/wake or Ctrl+Z/`fg` until the next keystroke or stream output; cursor landing mid-grapheme on Ctrl+E/A/K/U/arrow keys when an Indic conjunct or ZWJ emoji wraps across lines; vim operators corrupting text containing decomposed (NFD) accented characters; pasting text starting with `/` silently swallowing the input or triggering an unknown-command reply; pasting dumping stray escape sequences into the prompt when focus events or mouse-tracking reports interleave with bracketed paste; mouse wheel scrolling being too fast in Cursor and VS Code 1.92–1.104 due to an upstream xterm.js bug; scroll-wheel handling in JetBrains IDE 2025.2 terminals (spurious arrow keys, wrong-direction events, runaway acceleration); `/usage` Ctrl+S hanging when copying the stats screenshot to the clipboard on Linux/X11; `/terminal-setup` showing a contradictory error in Windows Terminal; `/effort` picker not reflecting the `CLAUDE_CODE_EFFORT_LEVEL` env var override; `/status` showing the wrong default model for some users; slash command autocomplete popup capped at ~3–5 visible commands instead of scaling with terminal height; statusline `context_window` token counts reflecting cumulative session totals instead of current context usage; Alt+T (thinking toggle) not working on macOS terminals without 'Option as Meta' enabled; dead keyboard input on Windows after re-opening a background session from `claude agents`; unbounded memory growth (10GB+ RSS) when a stdio MCP server writes non-protocol data to stdout; MCP servers that connect but fail `tools/list` silently showing 0 tools — they now retry once and show 'connected · tools fetch failed' in `/mcp`; unauthorized claude.ai MCP connectors showing as 'failed' instead of 'needs auth', and headless `-p` mode retrying non-transient 4xx connection failures; visual consistency improvements in slash command dialogs and `/login`, `/upgrade`, `/extra-usage` dialog spacing; `/tui fullscreen` startup banner updated to describe additional renderer benefits; Bedrock and Vertex 400 errors when `ENABLE_PROMPT_CACHING_1H` is set.

Wall Street Journal reports Anthropic inked a deal to use all compute capacity at SpaceX's Colossus 1 — Anthropic gains over 300 megawatts of capacity and over 220,000 NVIDIA GPUs at SpaceX's Memphis data center; the deal reflects Anthropic's strategy to seek diversified compute partners amid surging Claude demand, and runs alongside doubled Claude Code 5-hour rate limits for Pro/Max/Team/Enterprise customers all effective today, effectively resetting paid users' quota.

Reuters reports Anthropic unveils Dreaming AI agent self-improvement feature and Musk reverses stance on Anthropic — Dreaming (research preview) lets agents reflect on past conversations between sessions, curate memories, and surface recurring mistakes and team preferences, mirroring how the human brain consolidates memory during sleep; the same article notes that after SpaceX agreed to lease Colossus 1 capacity to Anthropic, Musk praised Anthropic on X for advancing 'AI for humanity', reversing his earlier criticism of Claude as 'evil' and saying he made the decision after spending time with Anthropic.

Anthropic held its Code w/ Claude 2026 developer conference in San Francisco — in his keynote, founder and CEO Amodei disclosed Q1 2026 revenue and usage grew 80x year-over-year (the company had planned for 10x), with ARR rising from $9 bn at the end of 2025 to $30+ bn and API usage on the Anthropic platform growing 17x year-over-year; the product strategy centers on maximizing capabilities of existing models rather than launching new ones, with no new model announced. Claude Code launched six new tools: Code Review, CI auto-fix (automated PR fixes), Security Reviews, Remote Agents, Routines (higher-order prompts that schedule background processes producing merge-ready PRs), and a Desktop app with async workflow support. Managed Agents simultaneously announced three new capabilities: Multiagent orchestration, Outcomes, and Dreaming (research preview). Speakers included CPO Ami Vora, Head of Product for Research Dianne Na Penn, Head of Product for Claude Code Cat Wu, and Claude Code creator Boris Cherny; full conference coverage in Simon Willison's live blog.

Anthropic launches three new features for Claude Managed Agents: (1) Multiagent orchestration in public beta — a lead agent divides complex tasks among specialist sub-agents, each with its own model, prompt, and tools, working in parallel on a shared filesystem and feeding context back to the lead agent, fully traceable through the Claude Console; (2) Outcomes in public beta — developers write a rubric describing what success looks like, and a separate grader evaluates the output in its own context window and prompts revisions; in internal testing, outcomes improved task success by up to 10 percentage points over a standard prompting loop with the largest gains on harder problems, and improved file generation quality by 8.4% for docx and 10.1% for pptx; (3) Dreaming in research preview — a scheduled process that reviews agent sessions and memory stores, extracts patterns, curates memories, and surfaces recurring mistakes, converged workflows, and team-wide preferences that a single agent can't see on its own; especially useful for long-running work and multi-agent orchestration. Memory also enters public beta.

Elevated errors across multiple Claude models — status.claude.com began investigating at 05-06 15:29 UTC, identified the issue and implemented a fix at 15:54 UTC, entered monitoring at 16:32 UTC, and resolved at 16:51 UTC; duration approximately 1 hour 22 minutes. Affected: claude.ai, Claude Console, Claude API, Claude Code, Claude Cowork.

Bloomberg reports Anthropic signed a computing deal with SpaceX to meet surging demand for its Claude AI software — Anthropic gains access to more than 300 megawatts of capacity and over 220,000 NVIDIA GPUs (H100, H200, and next-generation GB200) at SpaceX's Colossus 1 data center in Memphis within the month; the deal hands IPO-bound SpaceX a marquee customer to validate its AI ambitions to investors while easing Anthropic's capacity constraints amid surging demand for products like Claude Code; Anthropic also signaled interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity.

EPAM Systems announced a multi-year strategic partnership with Anthropic — EPAM is building a dedicated practice of more than 10,000 Claude-certified architects (including 250 specialized, forward-deployed engineer Black Belts) and has trained more than 20,000 EPAMers via Anthropic Academy; as of the announcement, over 1,300 architects are certified, with 5,000 expected by end of Q3 and thousands more in 2027; the partnership focuses on simplifying legacy operations, automating workflows, large-scale data analysis, and operational efficiency, combining EPAM engineering with Anthropic's Claude models, Claude Code, and Claude Agent SDK; the announcement coincides with Anthropic's Wall Street push of 10 financial services agent templates and the JPMorgan Chase strategic partnership.

Anthropic announces a compute partnership with SpaceX and raises Claude usage limits — gains the entire capacity of SpaceX's Colossus 1 data center in Memphis (more than 300 megawatts and over 220,000 NVIDIA GPUs including H100, H200, and next-generation GB200), coming online within the month; usage limits across paid Claude subscriptions rise correspondingly: Claude Code five-hour rate limits double within the month for Pro, Max, Team, and Enterprise customers, peak-time throttling for Pro and Max accounts is removed entirely, and API limits for Claude Opus models receive a major increase; Anthropic also expressed interest in partnering with SpaceX to develop multiple gigawatts of orbital AI compute capacity.

Claude Code 2.1.131 — Fixed the VS Code extension failing to activate on Windows due to a hardcoded build path in the bundled SDK (`createRequire` polyfill bug); fixed Mantle endpoint authentication failing with missing `x-api-key` header.

Claude Code 2.1.129 — Added `--plugin-url <url>` flag to fetch a plugin `.zip` archive from a URL for the current session; added `CLAUDE_CODE_FORCE_SYNC_OUTPUT=1` env var to force-enable synchronized output on terminals that auto-detection misses (e.g. Emacs `eat`); added `CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE`: when set on Homebrew or WinGet installations, Claude Code runs the upgrade command in the background and prompts to restart; plugin manifests' `themes` and `monitors` should now be declared under `"experimental": { ... }` (top-level still works but `claude plugin validate` warns); Gateway `/v1/models` discovery for the `/model` picker is now opt-in via `CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1` (was automatic in 2.1.126–2.1.128); Ctrl+R history picker now defaults to searching all prompts across all projects (matching pre-2.1.124 behavior), Ctrl+S to narrow to current project/session; third-party deployments (Bedrock, Vertex, Foundry, `ANTHROPIC_BASE_URL` gateway) no longer see spinner tips pointing at first-party Anthropic surfaces; `skillOverrides` setting now works: `off` hides from model and `/`, `user-invocable-only` hides from model only, `name-only` collapses description; the `claude_code.pull_request.count` OTel metric now counts PRs/MRs created via MCP tools, not just shell commands; policy refusal error messages now include the API Request ID for easier support debugging. Fixes: API errors with unrecognized 400 status codes showing raw JSON instead of the underlying error message; `/clear` not resetting the terminal tab title after a conversation; session title chip from `/rename` disappearing while a permission or other dialog is active; agent panel below the prompt being hidden when subagents are running (regression in 2.1.122); external-editor handoff (Ctrl+G) blanking the conversation history above the prompt; `/context` dumping its rendered ASCII visualization grid into the conversation, wasting ~1.6k tokens per call; `/agents` Library list arrow-key navigation: the highlighted agent now stays visible when the list exceeds the viewport; `/branch` success message not including the new branch's session id for `/resume`; bold headers with keycap/ZWJ/skin-tone emoji losing trailing characters in fullscreen mode; server-managed settings policy not applying for enterprise/team users whose stored OAuth credentials lacked the `user:inference` scope; OAuth refresh race after wake-from-sleep that could log out all running sessions; 1-hour prompt cache TTL being silently downgraded to 5 minutes; cache-miss warning appearing spuriously after `/clear` or compaction when changing `/effort` or `/model`; `Bash(mkdir *)`, `Bash(touch *)` and similar allow rules not being honored for in-project paths; `deniedMcpServers` patterns with a `*://` scheme wildcard not matching mixed-case hostnames; harmless WebSocket warning being logged as an error in `--debug` during voice mode; [VSCode] `/clear` not clearing the conversation context and displayed transcript.

Reuters relays a May 5 report from The Information that Anthropic has committed to spend roughly $200 billion with Google Cloud over five years for compute and chips. The commitment alone places Anthropic at more than 40% of Google's revenue backlog disclosed to investors; combined with OpenAI, Anthropic's and OpenAI's cloud contracts account for more than half of the roughly $2 trillion in long-term backlog held by the four major cloud providers (AWS, Microsoft Azure, Google Cloud, Oracle). Neither company has publicly confirmed the $200 billion figure; Reuters said it confirmed the contours of the deal but could not independently verify the dollar amount. The report sits alongside the same week's compute deals with SpaceX Colossus 1 (300 MW, 220K NVIDIA GPUs), AWS (5 GW, $100 billion), and Akamai (7-year, $1.8 billion), reflecting infrastructure pressure under Q1 annualized revenue and usage growing 80x year-over-year.

CNBC reports Anthropic founder and CEO Amodei warns of a 'moment of danger' — Mythos has uncovered tens of thousands of software vulnerabilities, leaving roughly a 6-to-12-month window before Chinese AI models catch up; an earlier Anthropic model found about 20 vulnerabilities in Firefox while Mythos found nearly 300; Amodei: 'The danger is just some enormous increase in the amount of vulnerabilities, in the amount of breaches, in the financial damage that's done from ransomware on schools, hospitals, not to mention banks', and 'this is about a moment of danger where if we respond to it correctly, and I think we started to take the first steps, then we can have a better world on the other side'; most Mythos-discovered vulnerabilities haven't been publicly disclosed because they remain unpatched and 'the bad guys will exploit' them; Amodei appeared alongside JPMorgan Chase CEO Jamie Dimon at Anthropic's New York financial services event, signaling Anthropic's lead in the enterprise AI market; Anthropic on the same day unveiled 10 banking and back-office AI agent templates and a unified Microsoft Office integration.

Fortune analyzes Anthropic's accelerating push into Wall Street with Moody's, Microsoft 365, and Jamie Dimon's endorsement — at an invite-only New York financial services briefing on 5/5, Anthropic launched 10 ready-to-run agent templates targeting the most labor-intensive workflows in banking, asset management, and insurance: pitchbook construction, KYC file screening, credit memos, underwriting, month-end close, statement audits, and insurance claims; three ecosystem deepenings were also announced: (1) full Microsoft 365 integration (Excel, PowerPoint, and Word add-ins generally available that day; Outlook in beta), with Claude functioning as a single agent that carries context across all four Office applications; (2) Moody's embeds its full platform into Claude as a native app, enabling analysis of credit ratings and risk data for over 600 million companies; (3) Claude Opus 4.7 shows substantial gains on financial work; through Dimon's joint appearance, the 5/4 JV with Goldman/Blackstone/H&F, and Microsoft 365 integration, Anthropic is building the 'operating layer for Wall Street'.

Anthropic launches Claude financial services agent templates and Microsoft 365 integration — debuts 10 ready-to-run financial services agent templates: Pitch builder (pitchbook construction), Meeting preparer, Earnings reviewer, Model builder, Market researcher, Valuation reviewer, General ledger reconciler (including NAV calculations against books of record), Month-end closer (close checklists, journal entries, close reports), Statement auditor (consistency, completeness, and audit-readiness checks), and KYC screener (entity files and escalation packages); templates run as a Claude Cowork or Claude Code plugin alongside the analyst, or as a Claude Managed Agent autonomously on the Claude Platform for batched deal books or nightly schedules; Microsoft 365 add-ins ship in parallel: Excel, PowerPoint, and Word generally available, Outlook coming in beta, with Claude operating as a single agent that carries context across all four Office applications; Moody's embeds its full platform as a native Claude app for analyzing credit ratings and risk data on over 600 million companies; the launch event in New York featured JPMorgan Chase CEO Jamie Dimon alongside Anthropic founder and CEO Amodei.

Anthropic Fellows, MATS, and Redwood Research jointly publish new research on sandbagging weak supervision — investigates whether a more capable model that strategically sandbags on tasks humans can't fully verify can be corrected when the only available supervision comes from weaker models; the paper finds that weak supervision can still train strategically sandbagging models back to near-full capability; output of the Anthropic-Redwood MATS stream, announced by co-author Emil Ryd on X on 5/5 and reposted by Anthropic on 5/6.

Claude Code 2.1.128 — 30+ bug fixes and several new features: bare `/color` picks a random session color; `/mcp` shows tool count for connected servers and flags servers with 0 tools; `--plugin-dir` accepts .zip plugin archives; `--channels` works with console auth (managed orgs require `channelsEnabled: true`); `/model` picker consolidates duplicate Opus 4.7 entries with current Opus shown as 'Opus'; OpenTelemetry: subprocesses (Bash, hooks, MCP, LSP) no longer inherit `OTEL_*` environment variables, preventing OTEL-instrumented apps from picking up the CLI's OTLP endpoint; fixes `CLAUDE_CODE_SHELL_PREFIX` argument corruption for MCP stdio servers when set with arguments containing spaces or metacharacters; SDK hosts receive persistent suggestions for Bash permission prompts via `localSettings`, written to `.claude/settings.local.json`; fixes crash loop when piping large input (>10MB) via stdin, MCP tool results dropping images with structured content, `/rename` failures on resumed sessions, tab navigation in `/config` stranding focus, parallel shell tool calls canceling on read-only failures, vim NORMAL-mode Space-key behavior, `/plugin update` not detecting npm-sourced plugin versions, stale `installed_plugins.json` entries polluting PATH, sub-agent summaries repeatedly firing and excessive token cost on idle agents, and multiple terminal rendering and hyperlink issues. Skipped v2.1.127 (never released).

Bloomberg reports the EU is in talks with Anthropic to get banks tested for Mythos-discovered flaws — Eurogroup finance ministers convened to discuss 'no European government currently has access to Mythos', with the White House having spent the last several weeks blocking Anthropic's plan to expand Mythos access to roughly 70 additional organizations and no European bank on the launch list; Germany's central bank publicly called on the EU to demand access to Mythos, arguing European banks cannot realistically test their own infrastructure against threats an adversary armed with the model would be capable of mounting; banks across the eurozone of all sizes have voiced concern about Mythos access; Mythos was unveiled on 4/7 and can identify and in principle exploit zero-day vulnerabilities in every major operating system and web browser, finding nearly 300 vulnerabilities in Firefox alone.

Fortune analysis: Anthropic JV is a frontal challenge to the Big Three consulting industry — headline directly states 'Anthropic takes shot at consulting industry'; the analysis notes the global software-to-services spend ratio is 1:6 (a multitrillion-dollar consulting market), giving AI-native firms a disruption opportunity; cites Sequoia partner Julien Bek's April thesis that 'the world's next great company won't sell software at all, but outcomes' (legal services, financial analysis, insurance processing delivered by AI but billed like consulting), and frames this JV as 'that thesis, capitalized and staffed'; the new firm is a standalone entity with Anthropic engineers embedded, mirroring Palantir's forward-deployment model, combining implementation capability with ownership of the underlying model to undercut traditional consultants; Blackstone President Jon Gray says the goal is to break down 'one of the most significant bottlenecks to enterprise AI adoption — the scarcity of engineers who can implement frontier AI systems at speed'; Anthropic CFO Krishna Rao: 'Enterprise demand for Claude is significantly outpacing any single delivery model. This new firm brings additional operating capability to the ecosystem'; PE is the natural beachhead: 85% of buyers now factor AI-enabled finance capabilities into company valuations.

Anthropic launches a $1.5B AI enterprise services firm with Goldman Sachs, Blackstone, and Hellman & Friedman — the three anchor partners each contribute about $300M, Goldman adds $150M, with additional backing from Apollo Global Management, General Atlantic, Leonard Green, Singapore's GIC, and Sequoia Capital; the new firm (still unnamed) aims to embed Claude into the core operations of mid-size businesses, starting with PE-portfolio companies and expanding to healthcare, manufacturing, financial services, retail, and real estate; positioned not as a traditional consultancy but as a venture deploying engineers to redesign workflows; Goldman global head of asset and wealth management Marc Nachmann said 'There's a big shortage of people who know how to apply these tools into businesses and then transform them' — combining technology with business context for real implementation; comes as both Anthropic and OpenAI prepare for IPOs as soon as this fall.

Elevated errors on Claude Opus 4.7 — status.claude.com identified the issue and implemented a fix at 05-04 14:07 UTC, resolved at 14:33 UTC, lasting approximately 26 minutes.

Axios comparative report: OpenAI announces a parallel $10B JV 'The Deployment Company' the same day, mirroring Anthropic's PE strategy — investors include TPG (anchor), Brookfield, Advent International, Bain Capital, Goanna Capital, and 14 others (19 total), with the PE consortium contributing roughly $4B over 5 years; OpenAI guarantees PE backers a 17.5% annual return over 5 years and adopts a Palantir-style forward-deployed engineer model; strategic intent matches Anthropic's JV — solving the AI-adoption choke point of 'how do enterprises use these tools in an ROI-positive way', with PE supplying portfolio companies as a built-in distribution channel focused on mid-sized firms; Anthropic's head of product for financial services Nicholas Lin told Axios 'There's a big gap between what AI can do today and the value the market is truly getting from it', noting the JV will help Anthropic scale customer support and develop deployment templates that copy across firms; demand is currently strongest for coding tools, expanding to finance and healthcare.

Elevated errors on Claude Opus 4.5 and Sonnet 4.5 — status.claude.com began investigating at 05-04 13:59 UTC, identified at 14:08 UTC, entered monitoring at 14:27 UTC, and resolved at 14:45 UTC, lasting approximately 46 minutes.

Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs announce a new AI-native enterprise services company — a standalone entity that embeds Anthropic Applied AI engineers directly inside customers to design and maintain Claude deployments, identify highest-impact use cases, and build custom solutions; total committed capital $1.5B with Anthropic, Blackstone, and Hellman & Friedman each anchoring at $300M and Goldman Sachs at $150M, plus General Atlantic, Leonard Green, Apollo Global Management, GIC, and Sequoia Capital filling out the consortium; the partners will use their own portfolio companies as the initial proving ground before targeting mid-sized companies, especially in PE-owned healthcare, manufacturing, financial services, retail, and real estate; OpenAI is reportedly pursuing a near-identical structure with TPG and Bain Capital; Anthropic is eyeing a public listing as soon as this year.

Elevated errors on Claude Opus 4.5 — status.claude.com began investigating at 05-04 08:09 UTC, identified at 08:19 UTC, continued working on the fix at 09:14 UTC, entered monitoring at 09:23 UTC, and resolved at 09:49 UTC, lasting approximately 1 hour 40 minutes.

The Department of War announces classified-network AI deployment agreements with eight technology companies — SpaceX, OpenAI, Google, NVIDIA, Microsoft, AWS, Reflection, and Oracle — pointedly excluding Anthropic; the exclusion stems from a supply-chain risk designation Defense Secretary Hegseth formalized in March after Anthropic refused to authorize Pentagon use of Claude for 'all lawful purposes' (Anthropic's stance: the language could enable domestic mass surveillance or fully autonomous weapons); during Congressional testimony Hegseth called Anthropic founder and CEO Amodei an 'ideological lunatic' and on X declared 'effective immediately, no contractor, supplier, or partner' doing business with the Pentagon may engage in any commercial activity with Anthropic; Anthropic has filed two suits in San Francisco and Washington, D.C., asking federal judges to overturn the order, with both cases ongoing; tensions eased somewhat after Trump Chief of Staff Wiles and senior officials met Amodei at the White House on 4/17, and Trump told CNBC a deal was 'possible'.

Fed Vice Chair for Supervision Bowman, in a speech at the FSOC AI Series Roundtable (theme: cybersecurity and risk management, held in Washington 4/27), used Anthropic's Mythos as an example to argue AI supervision must adapt dynamically — directly citing Mythos as evidence of the technology's 'dynamic nature' and noting its dual-use character: it lets firms patch self-identified vulnerabilities to enhance cybersecurity, but if used maliciously it could identify and exploit weaknesses; mentions that Treasury Secretary Bessent and Fed Chair Powell already convened major banks to discuss Mythos cybersecurity implications; on 4/17 the Fed/OCC/FDIC issued SR letter 26-2 narrowing existing model-risk-management guidance to traditional models and basic AI, leaving generative and agentic AI (including Mythos) outside the framework while regulators develop a new one.

On CNBC's 'Squawk Box', Pentagon CTO Emil Michael draws an explicit distinction — Anthropic itself remains on the supply-chain-risk blacklist, but its Mythos model 'is a separate national security moment'; direct quote from Michael: 'the Mythos issue that's being dealt with government-wide, not just at Department [of] War, is a separate national security moment where we have to make sure that our networks are hardened up, because that model has capabilities that are particular to finding cyber vulnerabilities and patching them'; he also stresses 'whatever guardrails, whatever principles they want to develop against has to be consistent with our values, our mandate, our restrictions even'; the interview airs the same day the Department of War excludes Anthropic from its eight-vendor classified-network AI list, echoing the softening signal Trump sent CNBC after Amodei's 4/17 White House meeting that a deal was 'possible'.

Claude Code 2.1.126 — Model picker now lists models from the gateway's `/v1/models` endpoint (when `ANTHROPIC_BASE_URL` points to an Anthropic-compatible gateway); adds `claude project purge [path]` to delete all Claude Code project state (transcripts, tasks, file history, config), with `--dry-run` / `-y/--yes` / `-i/--interactive` / `--all` flags; `claude auth login` now accepts pasting the OAuth code directly into the terminal for environments where browser callback cannot reach localhost (WSL2, SSH, containers); OpenTelemetry `claude_code.skill_activated` event adds `invocation_trigger` attribute (`user-slash` / `claude-proactive` / `nested-skill`); `--dangerously-skip-permissions` now bypasses prompts for writes to `.claude/`, `.git/`, `.vscode/`, shell config files, etc.; improved PowerShell 7 detection on Windows (Microsoft Store / MSI / .NET global tool), and PowerShell becomes the primary shell on Windows when enabled, replacing the default Bash; fixes auto-downscaling of images larger than 2000px, removal of oversized images from history, OAuth timeout on slow connections and IPv6-only devcontainers, stream idle timeout after Mac sleep and during long thinking pauses, Japanese/Korean/Chinese character rendering on Windows, `Ctrl+L` clearing prompt input, clipboard writes exposing content to EDR/SIEM telemetry, deferred tools unavailable in skills with `context: fork`, and several other issues; security fix: `allowManagedDomainsOnly` / `allowManagedReadPathsOnly` were ignored when a higher-priority managed-settings source lacked a `sandbox` block. v2.1.124 and v2.1.125 were skipped (never released).

Implicator.ai analysis: Claude Security public beta is Anthropic's strategic detour around Mythos access restrictions — the White House blocked Anthropic's plan to expand Mythos access to roughly 70 organizations, citing security and compute concerns; Claude Security gives Enterprise customers a supervised vulnerability-discovery path built on the safer Opus 4.7, sidestepping the federal review of Mythos; via service partners (Deloitte, Accenture, PwC), it reaches enterprise security business and converts revenue without depending on Mythos authorization; UK AI Security Institute scored Opus 4.7 at 48.6% on expert cyber tasks, behind GPT-5.5 (71.4%) and Mythos Preview (68.6%), but testing scenarios lacked active defenders and defensive tooling, limiting real-world applicability.

Anthropic Claude Security public beta — a code vulnerability scanning tool powered by Opus 4.7 with the model 'reasoning about code like a security researcher' by tracing component interactions and data flows across files (not just pattern matching); scans entire repos or specific directories/branches with each finding including confidence rating, severity, impact analysis, and reproduction steps, then generates targeted patch instructions that users can open in Claude Code on the Web to apply the fix in context; supports scheduled scans, CSV/Markdown export, Slack/Jira webhooks, and finding dismissals with documented audit reasons; Enterprise customers can enable it from the admin console, with Team and Max coming soon; technology partners: CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, TrendAI, Wiz; services partners: Accenture, BCG, Deloitte, Infosys, PwC.