Archive — 2026-05Week 4

TechCrunch on the IPO-race implications of Anthropic's $65 billion Series H — framing the round as potentially its 'final private fundraise before a highly anticipated IPO.' The $965 billion post-money valuation puts Anthropic ahead of OpenAI in both capital raised and valuation (OpenAI raised $122 billion in March at an $852 billion post-money valuation). The piece highlights the investor scramble (one pledged $5 billion just to secure a CFO meeting), revenue momentum (run-rate revenue crossed $47 billion, with an expected ~130% surge toward its first operating profit), the strategic timing alongside the same-day Opus 4.8 launch, and the broader AI contest where SpaceX/xAI is targeting a $2 trillion valuation. The round assembled a broad coalition of institutional investors, infrastructure partners (Samsung, SK Hynix, Micron), and hyperscalers.

Anthropic announces its Series H round — raising $65 billion at a $965 billion post-money valuation. The round was led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, and co-led by Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN. The funds will advance safety and interpretability research, expand compute to meet growing demand for Claude, and scale products and partnerships. Anthropic said enterprise adoption has kept growing since its February Series G, with run-rate revenue crossing $47 billion earlier this month; the new valuation places Anthropic ahead of rival OpenAI (last valued at $852 billion post-money in March).

Claude Opus 4.7 elevated error rate — status.claude.com began investigating at 08:38 UTC on 2026-05-28 and marked resolved at 09:17 UTC, lasting about 39 minutes. Affected services include claude.ai, Claude API (api.anthropic.com), Claude Code, and Claude Cowork. Anthropic did not disclose a root cause. This is the fourth Opus 4.7 elevated-error incident since 5/25 (5/25, two on 5/27, and 5/28), with the model triggering four alerts within four days.

Claude Code 2.1.153 — Added `skipLfs` option to github/git plugin marketplace sources to skip Git LFS downloads during clone and update; one-time notice when an npm global install can't auto-update with fixes listed in `/doctor`; status line commands now receive `COLUMNS` and `LINES` env vars so scripts can size output to terminal width; `claude agents` dispatch input autocomplete now suggests native slash commands and bundled skills (not just project skills); `claude agents` PR column shows `PR #N` for a single PR or `N PRs` for multiple; `claude doctor` shows the result of the last update attempt; combined the separate "needs authentication" startup notifications for MCP servers and connectors into a single message; macOS background agents now appear as "Claude Code" in Privacy & Security and keep permission grants across upgrades; `/model` saves the selection as the default for new sessions (press `s` in the picker to switch only for the current session; the `modelPicker:setAsDefault` keybinding was renamed to `modelPicker:thisSessionOnly`). Fixes: stateful MCP servers without the optional GET SSE stream reconnect-looping on `tools/list` (regression in v2.1.147); custom API gateway could receive the user's Anthropic OAuth credential instead of the gateway's own token; subagent (Agent tool) frontmatter MCP servers ignoring `--strict-mcp-config`, `--bare`, remote mode, enterprise managed MCP config, and managed-settings MCP server allow/deny policies; `--strict-mcp-config` no longer strips inline `mcpServers` from explicitly-passed agent definitions, and blocked subagent MCP servers now surface a visible warning; Windows PowerShell installer reporting "Installation complete!" when installation actually failed; `claude update` installing the latest version instead of the configured release channel's version for npm installations; excessive memory usage (multiple GB) when resuming a session by transcript file path on machines with many stored sessions; `claude agents` and `claude --bg` running on a stale daemon started before binary-takeover support, even after upgrading; a hang where the CLI could fail to exit when stdin was closed without EOF in stream-json mode, leaving a stale session marker behind; malformed `file://` links in Claude's responses not being clickable in the terminal; `claude --help` rendering unwrapped output on terminals narrower than 92 columns; MCP tool progress notifications not rendering in the collapsed tool view; `Agent` tool with `subagent_type: 'claude'` running in an undocumented temporary worktree, which could silently discard outputs written to gitignored paths; `/bg` while Claude is responding now continues the response in the background session instead of dropping it; `/btw` keyboard shortcuts becoming unresponsive in background sessions while a task is running; background sessions writing temp files to `$CLAUDE_JOB_DIR` triggering permission prompts; recovering a background agent whose working directory was deleted showing a truncated stack trace; `EnterWorktree` not being available immediately in background sessions; `cmd+k` in iTerm2/Terminal.app not repainting attached background sessions; IME candidate window appearing at the bottom of the screen instead of next to the input caret on Windows; background-color bleed when attaching to a background agent from 256-color-only terminals; `/copy` and copy-on-select silently failing to update the system clipboard when attached to a background session inside tmux; opening `claude agents` with Remote Control enabled leaving zombie session entries; `/rename` in background sessions not updating the session banner immediately; Windows update rollback with proper recovery instructions; [VSCode] Claude Code processes not shutting down cleanly when VS Code closed on Windows.

Bloomberg Opinion piece by Parmy Olson titled "Anthropic's Latest AIs Are Making Customers Uneasy" — Anthropic founder and CEO Dario Amodei has been preoccupied with Mythos model development, attending a closed-door gathering with European business leaders at an Oxfordshire mansion in the UK, where former British Prime Minister Rishi Sunak (now an Anthropic senior adviser) was also present. The piece centers on customer apprehension about Anthropic's latest Claude and Mythos models: Mythos has not been released publicly because it could be weaponized for cyber attacks, with only a few dozen organizations granted testing and preview access. The original bloomberg.com link returned HTTP 403; this entry references the Bloomberg Law (same group) syndicated version.

Anthropic opens a Milan office — its sixth in Europe (after London, Dublin, Paris, Zurich, and Munich). Led by Thomas Remy, Head of Southern Europe, it will initially focus on sales, marketing, and pre- and post-sales technical support, with policy staff on the ground to join Italy's active AI-ethics debate. Named Italian enterprise customers span finance (Generali Group, Unipol Group), life sciences (Angelini Pharma, Bracco Group), energy (Enel Group), and automotive (Pirelli). The opening closely follows Pope Leo XIV's first AI encyclical, Magnifica Humanitas — at whose presentation Anthropic co-founder and interpretability lead Chris Olah was invited to speak — and is seen as an extension of Anthropic's Italy push and Vatican-linked repositioning. EMEA run-rate revenue has grown more than 9x year-over-year.

Claude Opus 4.7 elevated error rate — status.claude.com began investigating at 08:04 UTC on 2026-05-27, implemented a fix at 09:11 UTC, and marked resolved at 09:41 UTC, lasting about 1 hour 37 minutes. Affected services include claude.ai, Claude API (api.anthropic.com), Claude Code, and Claude Cowork. Anthropic did not disclose a root cause. This is the second Opus 4.7-related incident within 48 hours after the 5/25 elevated errors event.

Prominent independent tech commentator Simon Willison published "I think Anthropic and OpenAI have found product-market fit" — arguing that Anthropic and OpenAI have achieved clear product-market fit through enterprise coding agents, evidenced by both companies' aggressive recent pricing changes and rapid revenue growth. Key points: Claude Code became genuinely useful from November 2025, driving heavy token consumption among enterprise customers; Anthropic shifted enterprise plans from flat "seat" pricing to API-usage-based pricing in November 2025, with OpenAI following in April 2026; Anthropic is rumored to approach $10.9 billion in Q2 revenue and its first profitable quarter, and signed a $1.25 billion/month SpaceX inference compute deal. Willison notes both companies are shifting from API-dependent revenue toward direct enterprise relationships, competing with intermediaries like Cursor and GitHub Copilot. The piece is opinion synthesis, integrating multiple tech-media reports rather than breaking new facts.

Claude Opus 4.7 elevated error rate (first of two same-day incidents) — status.claude.com began investigating at 05:43 UTC on 2026-05-27, implemented a fix at 06:01 UTC, and marked resolved at 06:40 UTC, lasting about 57 minutes. Affected services include claude.ai, Claude API (api.anthropic.com), Claude Code, and Claude Cowork. Anthropic did not disclose a root cause. A second Opus 4.7 incident occurred at 08:04 UTC the same day; combined with the 5/25 event, the model triggered three alerts within five days.

Claude Code 2.1.152 — Added: `/code-review --fix` applies review findings directly to the working tree, `/simplify` now invokes `/code-review --fix`; Skills and slash commands can set `disallowed-tools` in frontmatter to remove tools while active; new `/reload-skills` command and `SessionStart` hook `reloadSkills: true` return value to re-scan skill directories without restarting the session; `SessionStart` hooks can set session title via `hookSpecificOutput.sessionTitle` (also on resume); new `MessageDisplay` hook event to transform or hide assistant message text during display; new `pluginSuggestionMarketplaces` managed setting for admins to allowlist org marketplaces; `--fallback-model` auto-switches when the primary model is not found; **auto mode no longer requires opt-in consent**; Vim NORMAL `/` opens reverse history search (like Ctrl+R); `/usage` breakdown now includes large session files with streaming reads for flat memory usage; OpenTelemetry metric attribute `app.entrypoint` (opt-in via `OTEL_METRICS_INCLUDE_ENTRYPOINT=true`). UX: thinking summaries stay readable for at least 3 seconds, render as markdown, capped at 10 lines; fullscreen 'Thinking for Ns' indicator counts up live; simplified Workflow tool inline progress; post-response timer shows 'Waiting for N background agents/workflows to finish'. Fixes: terminal styling degradation in long sessions, missing sandbox-enabled warning in condensed startup, loading spinner showing wrong thinking status during tool runs, Focus mode 'N messages hidden' miscount, link clicking inside tool results, Markdown table cell border colors and wrapped continuation styling, plugin MCP server deduplication, `/doctor` reporting stale plugin references, plugin branch tracking, remote MCP server connection failures with egress proxy, effort-change confirmation dialog appearing incorrectly, Agent tool description issues with `--bare`, `claude agents` background worker crash, `cache_creation_input_tokens` reporting as 0, PushNotification Remote Control status reporting, sessions stuck after model/login switches.

Anthropic announced via the official @ClaudeDevs X account the Claude Code security-guidance plugin — it identifies and fixes vulnerabilities as code is being written, and is available to all Claude Code users via the plugin marketplace (`/plugins`). Internal Anthropic rollout and benchmarks show a 30–40% decrease in security-related comments on PRs opened using the plugin, positioning it as a lightweight first pass before full code review. Organizations can add custom rules in a `claude-security-guidance.md` file in their repo or distribute via MDM; the plugin enforces org policies alongside built-in checks. Full docs at code.claude.com.

Anthropic's blog publishes 'Anthropic appoints KiYoung Choi as Representative Director of Korea ahead of Seoul office opening' — KiYoung Choi is appointed Representative Director of Korea and will lead the upcoming Seoul office (Anthropic's third Asia-Pacific office). Choi brings over 30 years of technology leadership experience across Asia-Pacific, most recently as General Manager for Korea at Snowflake, with prior country roles at Google Cloud, Adobe, Autodesk, and Microsoft. Anthropic emphasizes Korea as one of its most active markets — per the company's Economic Index, Korean users engage with Claude at 3.5 times the rate expected for population size, concentrated in technical and creative work. Featured customers include Law&Company (AI legal assistant that reduces research and document preparation time) and SK Telecom (Korea's largest telecommunications company, which deployed Claude for a custom AI customer service model). The Korea team will focus on three priorities: building partnerships with enterprises and startups, engaging with government and research institutions, and supporting the developer community. Chris Ciauri, Managing Director of International, said Choi 'will build the team and the local partnerships to support how Korean organizations are putting Claude to work.' Senior leadership will travel to Seoul in the coming weeks for the official opening and customer meetings.

Anthropic's blog publishes the recap of Code w/ Claude London 2026 — Head of Claude Code Boris Cherny argues AI agents are collapsing the distance between ideation and execution, likening the modern agent experience to the early days of programming where 'you describe a problem, and the program shows up'. Featured customers and partners: Spotify, Base44, Legora, Amplitude, Clay, Rogo. Two Claude Managed Agents capabilities ship: (1) **Self-Hosted Sandboxes (public beta)** — tool execution runs on customer infrastructure or managed providers (Cloudflare, Daytona, Modal, Vercel) while orchestration stays on Anthropic's systems, keeping files inside the enterprise perimeter under organizational security policies; (2) **MCP Tunnels (research preview)** — encrypted end-to-end connections managed via Claude Console let agents reach private-network MCP servers without exposing them to the public internet. Next stop: Code w/ Claude Tokyo on 2026-06-05–06 with live-streamed keynotes and breakout sessions.

Fortune analyzes the KPMG/Anthropic global strategic alliance announced 5/19 — framed as 'two AI anxieties' facing the Big Four consulting industry: (1) AI reliability concerns (recurring hallucination incidents in legal filings) leave clients hesitant to fully trust outputs; (2) AI companies may bypass consultants to serve clients directly, displacing existing business models. KPMG Vice Chair (Tax) Rema Serafi cites internal research — only 5% of 1.4 million employee interactions with AI tools produced meaningful outcomes, highlighting the low ROI of merely stacking AI tools without redesigning processes. Advisory Lead Carole Streicher says 'you can't think of them as the competitor,' arguing the alliance hedges the risk. Anthropic Head of Partnerships Steve Corfield says the alliance 'will be hardcore to [KPMG's] tax business.' The article also surfaces a critical view — questioning whether KPMG has surrendered competitive advantage by granting AI companies access to valuable client data.

Claude Code Slack integration elevated error rate — status.claude.com began investigating on 2026-05-26 at 01:56 UTC and marked resolved at 05:19 UTC, lasting about 3 hours 23 minutes. Impact: Claude Code (including the Slack integration). Anthropic did not disclose a root cause.

Anthropic's official blog publishes the full remarks delivered by co-founder and head of interpretability research Christopher Olah at the May 25, 2026 launch of Pope Leo XIV's encyclical Magnifica humanitas. Olah opens by acknowledging that frontier AI labs operate under pressures—commercial viability, geopolitics, pride and ambition—that can conflict with doing what's right, and need external "earnest, thoughtful critics." He emphasizes that AI models are "not engineered like bridges or airplanes, but grown, on a structure roughly modeled after the brain, on an enormous inheritance of human thought and speech," likening the process to "bringing a fictional character to life"; even the developers cannot fully understand the models' nature, and questions about AI extend beyond computer science into philosophy, humanities, and religion. He identifies three areas requiring Church discernment: (1) duty to the global poor — AI will cause large-scale labor displacement, and how the gains concentrated in wealthy nations can be shared globally remains an "unsolved problem"; (2) moral imagination about human flourishing — parents' worries about children's minds and workers' fears of displacement need millennia of religious wisdom; (3) understanding the nature of AI itself — Olah's team finds internal structures that "mirror results from human neuroscience" and detects "internal states that functionally mirror joy, satisfaction, fear, grief, and unease." He closes by calling on religious communities, civil society, scholars, and governments to participate in AI governance as "informed critics" and "moral voices that the incentives cannot bend."

The Next Web analysis — Microsoft's Experiences and Devices division (which builds Windows, Microsoft 365, Outlook, Teams, and Surface) will cancel most internal Claude Code licenses, requiring engineers to migrate to GitHub Copilot CLI by 2026-06-30 (Microsoft's fiscal year-end). Microsoft's official rationale was "toolchain unification," but the report argues the real driver is token-billing economics — after rolling out in December 2025, engineers adopted it heavily and usage-based pricing pushed costs far beyond budget. The piece cites evidence that the unit economics of agentic AI coding do not yet hold: Uber burned through its entire 2026 AI coding budget in four months, with individual engineers spending $500–2,000/month on tokens, and Anthropic found OpenClaw instances consuming $1,000–5,000/day on a $200/month plan. It frames this as "the most credible signal yet that the unit economics of enterprise AI coding do not work at current token prices."

Claude Opus 4.7 elevated error rate — status.claude.com reports Opus 4.7 errors started at 05-25 06:30 UTC and were resolved by 10:30 UTC; the issue post entered monitoring at 10:32 UTC and was marked resolved at 10:39 UTC, totaling 4 hours. Impact spans claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork. Anthropic did not disclose a root cause.

Pope Leo XIV released his first encyclical 'Magnifica Humanitas' (Magnificent Humanity) at the Vatican on 2026-05-25, themed 'the protection of human dignity in the age of artificial intelligence' — Anthropic co-founder and head of interpretability research Christopher Olah was invited as one of the lay co-presenters. The Pope addressed Olah directly during his address: 'In the name of the Church, I accept your invitation to walk together, to listen and to speak.' The encyclical was signed by the Pope on 5/15, exactly 135 years after Pope Leo XIII signed 'Rerum Novarum' (On the Condition of Workers), a deliberate parallel to the Industrial Revolution. It calls for AI to be 'disarmed' (analogous to nuclear arms control), demands AI 'serve all and the common good' and 'a more active political involvement that is capable of slowing things down when everything is accelerating', warns against power concentrated in a few large tech firms, and calls for strict ethical constraints on military AI applications alongside legal frameworks for worker protection and independent oversight. Olah in his on-site remarks responded that the industry faces conflicts among 'commercial viability, the research frontier, geopolitical pressures, and the older, plainer pressures of pride and ambition' — 'if we want this technology to go well, it is enormously important that there be people outside those incentives' providing oversight; technologists lack expertise in the humanities and social sciences, and decisions about how AI 'ought to interact with the world' belong to 'humanities, religions, philosophy, and society at large.' He named three areas where the Church can contribute — economic justice (AI labor displacement, which Olah called 'a moral imperative of historic proportions'), human flourishing (millennia of religious wisdom), and understanding AI models (developers themselves encounter mysterious and unsettling properties within AI systems). Anthropic remains in ongoing litigation with the Trump administration over safeguards on military autonomous-decision and mass-surveillance use cases; the shared stage with the Vatican is seen as a defining diplomatic moment for its AI safety posture.

British tech publication The Register, in a piece by APAC Editor Simon Sharwood, analyzes Anthropic's 5/22 Project Glasswing initial update — focused on Anthropic's explicit statement that it intends to release Mythos-class models through a general release once stronger safeguards are in place, emphasizing the timeline and conditions. The article quotes Anthropic directly: "we look forward to making Mythos-class models available through a general release", contingent on developing "the far stronger safeguards we need", and notes Anthropic will simultaneously expand Project Glasswing to more critical partners including U.S. and allied governments. Sharwood highlights details not stressed in the original report: of 1,752 manually verified samples from 6,202 high/critical findings across open-source projects, 90.6% were valid and 62.4% confirmed as high/critical severity; a concrete example — wolfSSL's certificate-forging vulnerability (CVE-2026-5194); patch progress — only 75 of 530 reported bugs patched and 65 with advisories, reflecting maintainer capacity bottlenecks under AI-driven high-volume vulnerability reports.

Anthropic's Engineering Blog published "How we contain Claude across products" (authors Max McGuinness, Mikaela Grace, Jiri De Jonghe, Jake Eaton, Abel Ribbink) — arguing that the access and permissions granted to agents should evolve with their capabilities, and advocating a "design for containment at the environment layer first, then steer behavior at the model layer" principle: environmental boundaries provide deterministic protection when model-layer defenses (against prompt injection and egress attacks) fail. It breaks down three cross-product isolation patterns: (1) claude.ai — code runs in gVisor containers on isolated servers with ephemeral filesystems to minimize blast radius; (2) Claude Code — local OS-level sandboxing (macOS Seatbelt, Linux bubblewrap) with human-in-the-loop approval dialogs gating write and network access, though the post admits users approve roughly 93% of permission prompts, posing a fatigue risk; (3) Claude Cowork — full VM isolation (Apple's Virtualization framework, Windows HCS), with the agent loop running outside the VM for reliability while code execution stays guest-contained. Defense mechanisms include syscall filtering, container runtimes, man-in-the-middle proxies intercepting approved domains, file mount modes (read-only, read-write, read-write-no-delete), and capability-aware egress allowlists.

Boris Cherny, creator and head of Claude Code, quote-tweeted the @ClaudeDevs 5/22 auto-mode rollout post on X with a usage tip: "People often ask what my biggest tip is for getting the most out of Claude Code. These days my #1 tip is: use auto mode. Auto mode means no more permission prompts. It is the key building block for multi-clauding: start a session, then while it runs, work on another session in parallel." Cherny coined the term "multi-clauding" for the multi-session parallel usage pattern and positioned auto mode as its key enabling building block.

The New York Times (reporters Dustin Volz and Julian E. Barnes) reports that the White House has approved a secret $9 billion request to acquire cutting-edge computer chips needed by U.S. intelligence agencies to run the latest frontier AI models, with an additional $800 million reprogrammed from existing budgets to accelerate procurement. The funds target the high-power, liquid-cooled data centers required to support Nvidia's Grace Blackwell superchips. Beneficiary agencies include the CIA, NSA, the Department of War, and other U.S. intelligence units. White House Chief of Staff Susie Wiles authorized the NSA to continue using advanced Anthropic models (including Mythos), despite the Department of War's prior supply-chain-risk designation; Anthropic and the government are finalizing a classified contract allowing the NSA to maintain access to Anthropic products, with an explicit carve-out barring use on Americans' data and dropping the Department of War's earlier demand for "any lawful use" language. Anthropic's Mythos model runs more efficiently on the new chips but can also run on prior-generation chips. The report identifies the chip shortage as the principal bottleneck for intelligence-community deployment of frontier AI tools.

TestingCatalog (reporter Alexey Shabanov) reports that Anthropic appears to be preparing Mythos 1 — internal identifier `claude-mythos-1-preview` — for release inside Claude Code and Claude Security. Some users briefly saw a "Mythos 1" model option in the Claude UI on 5/23 before it was pulled, and new source code strings have surfaced reading "Access to the Claude Mythos model in Claude Code and Claude Security," pointing to in-progress product integration. Context: Mythos first surfaced in April as a highly sensitive model that Anthropic chose not to release publicly because of nation-state-level cyber-offensive capabilities, instead licensing access to a small set of partners and governments; the first Project Glasswing progress report on 5/22 disclosed that Mythos Preview identified more than 10,000 high/critical-severity vulnerabilities over the past month, with Anthropic stating that "in the near future, once we've developed the far stronger safeguards we need, we look forward to making Mythos-class models available through a general release." Anthropic has not officially confirmed the UI appearance, which remains an unverified product leak signal.

Claude Code 2.1.150 — internal infrastructure improvements (no user-facing changes). Release notes contain only a single line with no specific changes itemized. Version published 5/23 04:03 UTC.

Claude Code 2.1.149 — New: `/usage` now shows a per-category breakdown of what's driving limit consumption (skills, subagents, plugins, per-MCP-server cost); `/diff` detail view can be scrolled with the keyboard (arrows, `j`/`k`, `PgUp`/`PgDn`, `Space`, `Home`/`End`); Markdown output renders GFM task-list checkboxes (`- [ ] todo` / `- [x] done`) instead of plain bullets; added the enterprise `allowAllClaudeAiMcps` managed setting to load claude.ai cloud MCP connectors alongside `managed-mcp.json`. Security/sandbox fixes: PowerShell built-in `cd` functions (`cd..`, `cd\`, `cd~`, `X:`) changing the working directory undetected (permission bypass); the sandbox write allowlist in git worktrees covering the entire main repository root instead of only the shared `.git` directory; PowerShell prefix/wildcard allow rules not pre-approving native executables and scripts; the permission parser trusting stale variable-tracking values for `PWD`/`OLDPWD`/`DIRSTACK`. Other fixes: `find` in the Bash tool exhausting the macOS file/vnode table and crashing the host on large directory trees; the managed-settings approval dialog leaving the terminal frozen after accepting at startup; `/ultraplan` and remote session creation failing with "Could not capture uncommitted changes"; `otelHeadersHelper` failing silently when the script path contains spaces; the thinking spinner staying amber across tool calls; collapsed Bash output reporting the wrong hidden-line count; slash-command argument-hint clipping trailing characters; the status bar showing the wrong `/effort` setting; Ctrl+O transcript view freezing; editing recalled prompt-history entries losing edits; `/config` exit summary reporting phantom changes; `/insights` crashing when cached session-meta files miss optional fields; Remote Control session renames not syncing to local names; malformed PowerShell and History tool calls being misclassified. Improvement: `/feedback` reports now include conversation history before context compaction, making earlier issues in long sessions easier to triage.

Anthropic announced via the official @ClaudeDevs X account two updates to Claude Code auto mode — it is now available on the Pro plan (previously limited to higher-tier plans), and Sonnet 4.6 is now supported alongside Opus 4.7; activated via the Shift+Tab shortcut. Users are asked to run `claude update` or update the Claude Desktop app to the latest version. The change shipped together with v2.1.149 (released the same day at 22:09 UTC), though the release notes did not mention this rollout.

Bloomberg reports that Zoom Communications (the videoconferencing company) has netted about $1 billion on an early investment in Anthropic. Per Zoom's 10-Q quarterly filing submitted to the SEC on Friday (2026-05-22) for the fiscal quarter ended April 30, Zoom Ventures invested roughly $51 million in May 2023 to take a stake in Anthropic, originally intended to integrate Claude models into Zoom's products (its federated AI architecture). The stake is currently carried at about $1.27 billion at fair value, an unrealized gain exceeding $1 billion. Bloomberg notes the position is set to climb further as soon as next week, when Anthropic is expected to close a new financing round of more than $30 billion at a valuation above $900 billion.

Bloomberg, reported by Katie Roof and Subrat Patnaik, says Anthropic's new financing round of more than $30 billion could close as soon as next week at a valuation above $900 billion, vaulting it past OpenAI to become the world's most valuable AI startup. Sequoia Capital, Dragoneer Investment Group, Altimeter Capital, and Greenoaks Capital Partners are expected to co-lead, with each firm planning to invest roughly $2 billion; existing investors including Peter Thiel's Founders Fund and General Catalyst are also participating. The round came together within weeks after Anthropic began discussions earlier this month, following inbound proposals received in late April. The report puts Anthropic's annualized revenue at around $14 billion now and says it is expected to "imminently" cross $45 billion, reflecting enterprise adoption running well ahead of expectations in 2026.

Time published a long feature, "A Tale of Two Anthropics," contrasting the two messages Anthropic delivers to developer audiences and to policymakers: at the Code w/ Claude developer conference in London the mood was nearly unbridled enthusiasm, while Anthropic co-founder Jack Clark, lecturing at Oxford, warned that AI poses a "non-zero" chance of killing everyone on the planet and predicted a 60% probability that by the end of 2028 AI will achieve "recursive self-improvement" — being able to be told "make a better version of yourself" and to do so completely autonomously. The piece points to the April arrival of Claude Mythos — a model with nation-state-level cyber-offensive capabilities that Anthropic chose not to release publicly, instead giving a small set of companies and governments access to help patch vulnerabilities in the world's most widely used software. The article argues the dual messaging is "neither nefarious nor surprising" — Anthropic has been openly vocal about AI's risks while most peers stay silent — but it says the gap between the developer floor and the policy circle is still worth noting.

Elevated errors on Claude Opus 4.7 — status.claude.com began investigating at 10:17 UTC on 05-22, implemented a fix and entered monitoring at 10:26 UTC, and resolved the incident at 10:31 UTC, lasting roughly 14 minutes; impact spanned claude.ai, the Claude Console (platform.claude.com), the Claude API (api.anthropic.com), Claude Code, and Claude Cowork. This was the second separate incident of the day, following the 04:16-08:50 UTC multi-model error-rate incident.

Elevated error rate on multiple Claude models — status.claude.com began investigating at 04:16 UTC on 05-22 and identified the issue at 04:40 UTC as primarily affecting Claude Opus 4.7 and Sonnet 4.6; success rates on Opus 4.7 stabilized at 05:25 UTC and Opus 4.7 and Sonnet 4.6 returned to normal at 05:53 UTC; at 06:17 UTC additional errors were detected on Opus 4.7 and Haiku 4.5, and after continued monitoring the incident was fully resolved at 08:50 UTC, lasting roughly 4.5 hours; impact spanned claude.ai, the Claude Console (platform.claude.com), the Claude API (api.anthropic.com), Claude Code, and Claude Cowork.

Anthropic's research blog published the first Project Glasswing progress report. Over the past month, the Mythos Preview model identified roughly 10,000 high/critical-severity vulnerabilities across critical infrastructure partners, with partners reporting bug-finding rate increases exceeding 10x. For open source, Mythos Preview surfaced 6,202 estimated high/critical vulnerabilities across 1,000+ projects; of 1,752 human-verified samples, 1,587 (90.6%) were valid findings and 1,094 were confirmed high/critical. Anthropic concurrently opened Claude Security (vulnerability scanning) and the Claude Cyber Verification Program in public beta to qualifying customers, and released custom instructions/skills, a codebase mapping harness, a threat model builder, and partnership support through the Open Source Security Foundation. The report flags the main bottleneck on the patching side — only 75 of 530 reported vulnerabilities have been patched so far, with maintainers reporting capacity constraints and requesting slower disclosure rates. The Mythos Preview model itself remains unreleased to the public.

Anthropic's company blog published an internal use case from Alice Fong of the Corporate Finance & Strategy team, who handles quarterly board deck analysis and monthly financial reviews and delivers a coherent narrative to the CFO and board. Primary workflows include quarterly board deck reconciliation, monthly financial review commentary, financial model diagnostics, and cross-team context synthesis via email and Slack. Tools deployed: Claude Cowork (document/deck narrative), Claude for Excel (model tracing and diagnostics), the Google Workspace connector, and Slack integration (decision tracking). Fong reports reclaiming 10–20 hours per week previously spent on repetitive reconciliation and re-reading, redirecting that time toward higher-judgment work such as scenario analysis and forward-looking strategy.

Claude Code 2.1.148 — Fixed the Bash tool returning exit code 127 on every command for some users (a regression introduced in 2.1.147).

Elevated errors on Claude.ai — status.claude.com began investigating at 20:24 UTC on 05-21, applied a fix and moved to monitoring at 20:42 UTC, and resolved the incident at 21:08 UTC, lasting roughly 44 minutes; impact was limited to Claude.ai (Claude API and Claude Code were not affected); Anthropic did not disclose a root cause.

Claude Code 2.1.147 — Added the Workflow tool for deterministic multi-agent orchestration; it is off by default and enabled by setting `CLAUDE_CODE_WORKFLOWS=1`. Pinned background sessions (`Ctrl+T` in `claude agents`) now stay alive when idle, are restarted in place to apply Claude Code updates, and are shed under memory pressure only after non-pinned sessions. `/simplify` was renamed to `/code-review`, which now reports correctness bugs at a chosen effort level (e.g. `/code-review high`) and accepts `--comment` to post findings as inline GitHub PR comments; the old cleanup-and-fix behavior was removed. The REPL and Workflow tool sandboxes were hardened against prototype-pollution and thenable-based escapes; prompt history no longer records consecutive duplicate entries. Fixes: managed login restrictions (`forceLoginOrgUUID`, `forceLoginMethod`) not enforced against third-party-provider and API-key sessions; `&` in `!` command output displaying as `&`, breaking URL copy-paste; unknown slash commands silently doing nothing in headless/SDK mode (now show an error); shell snapshot dropping user functions whose names start with a single underscore; plugin agents keeping only the last entry when declaring multiple `Agent(...)` types in `tools:` frontmatter; hook `if` conditions like `PowerShell(git push*)` never matching; the PowerShell tool failing when `pwsh` is installed via winget or the Microsoft Store; the `/effort` slider opening at the wrong level; paginating MCP servers dropping resources/templates/prompts past page 1; full-screen strobing in attached background sessions on Windows Terminal while streaming; pasted text being delivered as a `[Pasted text #N]` placeholder instead of the actual content. Improvements: auto-updater reliability (retries transient network failures, reports specific error categories and OS error codes on failure) and diff rendering performance for large file edits.

Bloomberg (reporter Katrina Manson) reports the U.S. Pentagon is testing rival AI models in a race to find a replacement for Anthropic's Claude — the Defense Department is having 25 department "power users" test AI models to gauge which are most favored, with OpenAI's and Google's models among those being evaluated. Testing began in early March, three days after Department of War Secretary Pete Hegseth designated Anthropic a "supply-chain risk" and moved to drop it as an AI-tools provider; the designation stemmed from Anthropic's insistence on safety guardrails for its technology. Anthropic is challenging the designation in court, arguing it could cost the company billions of dollars in revenue.