Today on The Anvil: the design-to-code handoff race has three entrants in 72 hours — Anthropic ships Claude Design with a machine-readable bundle into Claude Code, just after Figma's agent went to closed beta and Google Stitch launched its streaming agent. WebMCP heads to a Chrome origin trial June 2 with 80% token-reduction demos already public. And a 7,000-gallon aerospace chemical tank in Garden Grove forces a 40,000-resident evacuation.
Anthropic launched Claude Design, a canvas-based prototyping tool that grounds in the codebase's actual design system and exports a structured handoff bundle that Claude Code consumes directly — no copy-paste, no DOM rebuild, no vendor sandbox. The pair runs in the same model family with the same codebase context. This is the supply-side answer to the bidirectional sync thesis Figma declared in March: rather than pulling production code into Figma as editable frames, Anthropic is collapsing the design and coding context into one conversation from the start.
Why it matters
Figma's agent (closed beta, paid plans), Google Stitch's streaming agent, and now Claude Design all shipped within 72 hours of each other — and they're attacking the same translation problem from opposite directions. Figma and Stitch are extending the canvas toward code; Claude Design extends the coding context toward canvas. The question that didn't exist a week ago: does the handoff bundle format converge (DESIGN.md-style open standard) or does each lab try to own its own schema? The answer will determine whether design system authors ship for one platform or three.
Three coordinated and unrelated supply-chain attacks landed inside a week. Megalodon injected malicious GitHub Actions workflows into 5,561 repos on May 18–19, harvesting CI/CD secrets and OIDC tokens. A poisoned Nx Console VS Code extension — live in the Visual Studio Marketplace for 18 minutes — stole 1Password/npm/GitHub/AWS creds from developers and pivoted into ~3,800 of GitHub's internal repos. And on May 22–23, Laravel Lang shipped Composer-autoload backdoors across 700+ historical package versions with a multi-stage stealer targeting 17 categories of secrets.
Why it matters
Last week's TanStack/SLSA-provenance attack already showed signed artifacts aren't a guarantee. This week's pattern is worse: the attack surface is the developer's IDE extension, the CI workflow file, and the framework's autoloader — all of which run before any human review. For a product team using Claude Code, Cursor, or any MCP-driven workflow, the practical move is to treat GitHub Actions YAML as production code, restrict OIDC token scopes, and audit every auto-updating extension's permissions. Marketplace review gates are no longer a meaningful defense.
Alibaba's Qwen team announced Qwen3.7-Max on May 20 — a 1M-token reasoning agent. In a benchmark disclosed this week, the model ran for 35 hours unattended, optimizing a hardware attention kernel for Alibaba's T-Head-ZW-M890 accelerator and achieving 10× speedup through 432 kernel tests and 1,158 tool calls. It scores 56.6 on the Artificial Analysis Intelligence Index (5th overall), up 4.8 points from its predecessor.
Why it matters
This is the cleanest published demonstration of a long-horizon autonomous coding run on an unfamiliar hardware target — not a toy benchmark, an actual kernel against an internal ASIC. The relevant detail for builders isn't the 10× number; it's that the model self-recovered from compilation errors and ran measurement loops without human intervention. The model is API-only and Chinese-origin, which matters for procurement, but the architectural pattern (long-horizon agent + tight compile/measure feedback) is what other labs will copy.
One month after Anthropic loosened Glasswing disclosure rules to let its 40–50 partners share Mythos findings externally, the program has surfaced 10,000+ high- or critical-severity vulnerabilities. Cloudflare alone found 2,000; Mozilla found 10× more Firefox bugs than previous testing rounds. Anthropic is now publicly warning that discovery rate has outpaced patch capacity — the widening window of known-but-unfixed exposure is the new problem.
Why it matters
The Glasswing loosening was framed last week as a governance improvement — responsible disclosure norms replacing internal siloing. This week's numbers reveal the downstream consequence: the pipeline is now generating more confirmed critical findings than the remediation ecosystem can absorb. The 90-day coordinated-disclosure norm was calibrated for human-paced discovery. Mythos has broken that assumption, and adversaries running equivalent capabilities are on the same curve.
A faulty 7,000-gallon methyl methacrylate tank at GKN Aerospace in Garden Grove is releasing toxic vapors and at risk of catastrophic failure. Incident Commander Craig Covey publicly framed it as binary — leak or blow up. Roughly 40,000 residents are evacuated across North Orange County, OCFA is coordinating nationwide hazmat expertise, and evacuation centers and air monitoring are active.
Why it matters
This is the largest industrial hazmat event in OC in years and a stress test of the county's emergency-response coordination two weeks before the June 2 primary. The aerospace cluster in northern OC sits inside dense mixed residential-industrial zoning, so a single tank failure cascades across multiple cities. Worth watching: how the post-incident review interacts with the Supervisor District races, several of which are already running on industrial safety and accountability themes.
The Laguna Beach City Council voted 3-2 to deny demolition of the Egan Homestead — believed to be the last remaining homestead property in Orange County (1909, originally granted under the Homestead Act of 1910) — overturning a September 2025 approval after months of historic-consultant review. Meanwhile, Dana Point officials provided a May 15 Coffee Chat update on the harbor revitalization: commercial-core demolition is complete, Mother's Market, Nectarine Grove, and Harwood House are confirmed tenants, the hotel naming contest is open, and a new Community Development Director (Jonathan Lightfoot) is managing parking, traffic and property maintenance during the build.
Why it matters
Two opposite OC coastal development outcomes the same week — Laguna chose preservation over a 40-acre development play, Dana Point chose phased modernization with active community programming during construction. The Egan reversal is a precedent that will be cited in future coastal-development hearings; the Dana Point cadence is a template for how to keep tenants and residents engaged through a multi-year harbor rebuild without losing the foot traffic that pays for it.
Anthropic shipped Claude Code v2.1.150 with per-category usage breakdown (/usage now shows skills, subagents, plugins, and per-MCP-server cost), keyboard-navigable detail view, sandbox write-allowlist corrections for git worktrees, PowerShell permission fixes, and Enterprise-managed cloud MCP connectors via managed-mcp.json. Twenty-plus bug fixes including stale variable tracking in bash cd/pushd/popd and macOS vnode-table exhaustion during recursive find(). This is the first release since the doubled rate limits and lifted Opus throttling that shipped alongside the SpaceX compute deal.
Why it matters
The Enterprise managed-MCP support is the strategically new piece: centrally controlled allow-lists for which cloud MCP servers agents can call is the operational governance primitive that's been missing since Anthropic opened the MCP write path. The NSA's MCP threat model (serialization vulnerabilities, trust boundary issues) dropped the same week — the managed-mcp.json mechanism is the deployment-level control the NSA document was implicitly asking for.
Volvo Autonomous Solutions and DSV started commercial depot-to-depot autonomous freight on the Dallas–Houston lane this week using the Volvo VNL Autonomous with Aurora Driver — safety driver still aboard, with expansion planned. Separately, the House released the BUILD America 250 Act on May 17, directing DOT to write a federal performance-based safety standard for automated driving systems in commercial vehicles within two years. No AV trucks operate under the new framework until the rule publishes.
Why it matters
Autonomous freight just moved from pilot to integrated commercial lane on one of the highest-volume US corridors, and Congress is finally moving to replace the 50-state regulatory patchwork. Both events matter together: the technology is reaching commercial maturity at the same moment the federal framework is being drafted, and the rulemaking will decide hours-of-service, pre-trip inspection, liability, and CSA scoring under a no-driver model. Anyone modeling freight cost curves for the back half of the decade needs to track this rulemaking.
Walmart has reached ~50% automation in U.S. e-commerce fulfillment centers: 60%+ of stores receiving freight from automated DCs, 26% global e-commerce growth, 36%+ of store-fulfilled orders delivered under three hours, and the Sparky agent showing 35% higher AOV and 100%+ weekly active user growth. Separately, McKesson posted FY2026 revenues of $403B (+12% YoY) while operationalizing AI orchestration across its North American pharma distribution network — a new Montreal robotic DC, integrated demand/supply/inventory/ops planning, and 120 additional patients per employee as the disclosed productivity benchmark.
Why it matters
Gartner called out 'agent washing' and the GEP/Darden study put 95% of supply chain AI pilots as failing to scale — both this week. Walmart and McKesson are the same-week production counterexamples, publishing operational KPIs rather than pilot results. The pattern matches Conagra's ICON 2026 disclosure (governance and process redesign first, agents second) and stands in direct contrast to Starbucks' nine-month AI inventory failure. The emerging empirical split: tier-1 operators with existing data infrastructure and process discipline are scaling; everyone else is in the 95%.
ModelRift published a parametric CAD benchmark using the Pantheon as a reference task, pitting Antigravity 2.0, Cursor, Claude Code, Codex Desktop and others against the problem of translating an architectural reference into clean OpenSCAD with proper geometry, symmetry, and export quality. Antigravity 2.0 won — but the more interesting finding is that the client UI and interaction speed mattered nearly as much as the underlying model, and the gap between 'preview looks right' and 'exports a clean parametric file you can manufacture from' remains substantial.
Why it matters
This is the cleanest production-grounded benchmark of AI in parametric CAD this year — ModelRift is routing real 3D geometry jobs based on the results, not running a leaderboard. For anyone working at the digital/physical seam, the takeaway is that frontier models alone don't close the gap to manufacturable output; the interaction layer (how fast you can iterate, how the agent shows intermediate states) is the actual differentiator. Expect parametric CAD to be the next domain where harness engineering eats prompt engineering.
Following the Chrome 149 announcement at Google I/O, the WebMCP origin trial opens June 2. New this cycle: VEKTOR Memory shipped seven WebMCP tools and reports ~80% token reduction versus DOM scraping, and a tutorial published this week walks through building an agent-ready product card with navigator.modelContext.registerTool. Microsoft Edge remains on board (~70% combined browser share); Firefox and Safari haven't committed.
Why it matters
The 'agent-readiness stack' thesis from earlier this week now has running demos and concrete cost data. The implication for frontend work is real: semantic HTML, accessibility trees, and component clarity stop being a11y nice-to-haves and become foundational agent-usability infrastructure. The interesting question is whether design system authors start shipping WebMCP tool surfaces alongside ARIA roles by default — that's where this becomes a component-library standard rather than a website-level API.
Two opposing infrastructure stories surfaced the same day. The Trump 2027 budget would eliminate the Community Development Block Grant program, which has sent Spokane $7.5M+ over the last two years for home repair, job training, and neighborhood projects. Meanwhile Garco|QD Builders is breaking ground on a $195M Central Hall at Spokane International Airport, and DEQ awarded Sandpoint $40.5M ($38M loan, $2.5M grant) for wastewater plant upgrades using an alternative 'wet weather treatment' approach that saves $10M over a conventional build.
Why it matters
Federal-to-local funding is decoupling: large state and federal capital programs are still flowing (airport, wastewater, INL grid R&D), but the discretionary block-grant layer that funds neighborhood-scale resilience is on the chopping block. For Spokane, the gap shows up first in SNAP-style nonprofits and low-income housing assistance, not headline projects. Worth watching how city council and the SPS levy timing decision interact with a possible CDBG cliff in FY27.
Pakistan's Field Marshal Asim Munir conducted a second rapid Tehran shuttle, and Al Hadath reports Iran and Pakistan have drafted a ceasefire explicitly excluding nuclear enrichment and ballistic missiles — terms the White House has already publicly rejected. Trump cancelled personal plans to convene national-security leadership and has reviewed strike-target packages. Rubio described talks as showing 'slight progress' while calling Iran's Hormuz tolling proposal 'unfeasible.' U.S. intel reportedly assesses Iran has restored 90% of its underground missile storage. The 12-vs-20-year enrichment gap and the Supreme Leader's ban on exporting the ~441 kg HEU stockpile — both hardened in the last 48 hours — remain the structural sticking points.
Why it matters
The leaked draft's structure makes the negotiating geometry explicit in a way it hasn't been before: Iran will accept a wide ceasefire on everything except nuclear and missiles, and the U.S. won't accept a ceasefire without those terms. That's not a gap that shuttle diplomacy closes; it requires one side to move a red line. The simultaneous strike-planning and Pakistan mediation tracks mean markets are pricing a bimodal outcome — deal or escalation — with narrowing middle ground.
A ruling in Chatrie v. United States is expected before end of June 2026 — the Supreme Court's first major digital-privacy case since Carpenter (2018). The case directly addresses geofence warrants, and the precedent could extend to reverse keyword searches and reverse AI-chat-history searches, where law enforcement queries population-scale data without naming a suspect. Documented cases already show geofence warrants misidentifying innocent people and exposing protest participants.
Why it matters
Whatever the Court decides, it sets the constitutional ceiling on the kind of dragnet techniques that have become routine in federal investigations and that increasingly run against AI chat logs and assistant query histories. For builders shipping any product that retains user prompts, location, or search histories, the ruling will likely shape both data-retention defaults and the legal posture for responding to law-enforcement requests. Worth flagging now so legal teams aren't scrambling in July.
Handoff geometry is the new product surface Claude Design → Claude Code, Figma → DESIGN.md, Stitch's streaming agent, and WebMCP all attack the same problem: lossy translation between layers. The differentiator is no longer the model; it's the shape of the bundle that crosses a boundary.
Supply chain attacks have crossed an inflection Megalodon's 5,561-repo GitHub Actions injection, the Nx Console marketplace compromise that pivoted into 3,800 GitHub internal repos, and the Laravel Lang 700-version backdoor all landed in the same week. The shared pattern: developer tooling and CI/CD secrets are the new soft target, with marketplace auto-update as the propagation vector.
Local infrastructure stress is the quiet headline Garden Grove's hazmat evacuation, the Spokane CDBG funding cliff, Sandpoint's $40.5M wastewater package, and the $195M Spokane airport hall all surfaced today. Federal-to-local funding turbulence is reshaping municipal capital plans across both of Clark's home regions.
Agentic platforms are absorbing the IDE Antigravity 2.0 ships as a standalone agent-first desktop app; Google's Managed Agents API turns a Linux sandbox into a single API call; Claude Code v2.1.150 adds enterprise MCP plumbing. The editor is no longer the center of gravity.
Iran talks at maximum ambiguity Day 85 brings simultaneous Pakistani shuttle diplomacy, a leaked Iran-Pakistan draft excluding nuclear and missile issues, and White House strike-planning sessions. HEU stockpile and Hormuz tolling remain the two unresolved hinges; the diplomatic and military tracks are now running in parallel rather than sequence.
What to Expect
2026-05-28—Huntington Beach deadline to submit compliant housing element or escalate to $50K/month penalties June 1; KCRCC contested chair election same day.
2026-06-18—Google ends free Gemini IDE access, pushing developers to paid Antigravity.
End of June 2026—Supreme Court expected to rule on Chatrie v. United States, setting geofence-warrant and reverse-search precedent.
Next 1–2 weeks—Iran negotiations: Pakistani mediation continues amid White House strike-planning; HEU and Hormuz tolling remain unresolved sticking points.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
958
📖
Read in full
Every article opened, read, and evaluated
167
⭐
Published today
Ranked by importance and verified across sources
14
— The Anvil
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste