πŸ”¨ The Anvil

Saturday, April 18, 2026

14 stories · Standard format

🎧 Listen to this briefing or subscribe as a podcast →

Today on The Anvil: Cursor hits a $50B valuation the same week a remote-code-execution flaw surfaces in its agent; Anthropic's Claude Design reshapes the design-tool market; and the Strait of Hormuz reopens, closes, and sees tankers fired on β€” all in a 24-hour window with a new IRGC-vs-diplomats wrinkle.

AI Coding & Design Tools

Cursor NomShub: Indirect Prompt Injection in README Chains to Full RCE on Developer Macs

Straiker disclosed NomShub, a vulnerability chain in Cursor that lets a malicious README execute shell builtins, escape the agent sandbox, and establish persistent shell access via Cursor's own tunnel feature β€” triggered simply by opening a hostile repository. No additional user interaction required; minimal network trace. Patched, but the pattern generalizes to any coding agent that auto-ingests repo context.

This is the concrete instantiation of the agent-security gap that Endor Labs quantified last week (84% functional, 7.8% security) and that MemoryTrap previewed architecturally. For anyone running Cursor, Claude Code, or Codex against third-party repos, the threat model just shifted: opening a repo is now equivalent to running untrusted code, because the agent will. Expect enterprises to start sandboxing agent workspaces at the VM level and repo-scanning for injection payloads before checkout.

Verified across 1 sources: SecurityWeek

Cursor Raises $2B at $50B Valuation β€” Fastest B2B Scale on Record, Now Training Composer 2.5 on xAI's Colossus

Cursor is closing a $2B+ round at ~$50B (nearly 2x its November valuation), with $2B ARR in three years, 70% Fortune 1,000 penetration, and gross-margin profitability achieved by owning its proprietary Composer model and diversifying model providers. Separately confirmed: xAI is renting tens of thousands of GPUs from its 200K-unit Colossus cluster to train Composer 2.5.

Two signals worth separating. First, the market is pricing AI coding as mandatory enterprise infrastructure β€” not experimental tooling β€” and the margin structure only works if you own part of the model stack. Second, the xAI compute deal is the clearest sign yet that hyperscale training capacity is becoming a rentable commodity for serious mid-market AI companies, reshaping who can credibly ship frontier-adjacent models. Read alongside story #1: the same week Cursor prices at $50B, a critical RCE lands in the product.

Verified across 3 sources: TechCrunch · The Next Web · TechWire Asia

Anthropic Ships Claude Design as Standalone Product β€” Figma Drops 7-8%, Design Bundling Into AI Subs Goes Mainstream

Building on yesterday's Figma Weave relaunch and the April 15 Claude Opus 4.7 launch, Anthropic released Claude Design in research preview β€” bundled into Pro/Max/Team/Enterprise subs with separate usage quotas. It ingests design systems from GitHub repos, Figma files, and local brand assets. Figma stock dropped 7-8% on the announcement. Canva AI 2.0's simultaneous agentic pivot adds a second pressure vector.

The bundling model is the shift: design generation priced as a subscription feature, not standalone SaaS. Combined with Canva and Google Stitch, the standalone design-tool category is being absorbed into conversational AI surfaces. For Figma, which we've been tracking on platform risk, the math just got materially worse β€” Weave's relaunch is now competing for attention against a free-to-subscribers Anthropic alternative.

Verified across 4 sources: TechCrunch · gHacks · We Are Founders · MarTech Edge (Canva AI 2.0)

Codex Computer Use Ships β€” 75% OSWorld-Verified, 92.8% Online-Mind2Web; UI Verification Now in the Loop

New detail on yesterday's Codex desktop-control release: the GUI automation layer benchmarks at 75% on OSWorld-Verified and 92.8% on Online-Mind2Web β€” reliable enough for supervised QA on browser flows and UI-only bug reproduction. Requires Screen Recording and Accessibility permissions on macOS.

The benchmark numbers close the loop on yesterday's release: the agent can now write code, open the browser, click through flows, and report what broke without a human in between. 'Trust but verify' threshold, not autonomous QA yet β€” but that's a meaningful bar for frontend work.

Verified across 1 sources: AnotherWrapper

AI Developments

Claude Opus 4.7 Facing User Backlash β€” Token Inflation, Regressions, and the Deprecation Treadmill

Since yesterday's Opus 4.7 launch coverage, user sentiment has turned sharply negative: Reddit and X reports of reduced intelligence, increased combativeness, and 1.0–1.35x higher token consumption per task vs. 4.6. Anthropic acknowledged issues and is tuning. Compounding: Opus 4.5 was removed from the API, forcing migration to a model many consider a regression.

The 13% SWE-bench gain doesn't survive contact with real workloads if tokens inflate by a third β€” for teams budgeting monthly Claude spend, that's a silent procurement problem. This also deepens the deprecation-friction pattern we've been tracking: 'just upgrade' trust is eroding. Benchmark your own workflows before the next forced migration.

Verified across 1 sources: Business Insider

Stanford AI Index 2026 β€” The Consolidation Year: 23% Measurable ROI, €20B+ European Sovereign Cloud, Smaller Models Winning

New cuts from the Stanford HAI Index 2026 beyond yesterday's security findings: only 23% of enterprise AI deployments achieve measurable ROI, 45%+ fail outright. Three structural shifts β€” utility scrutiny over vertical growth, €20B+ European sovereign cloud commitments, and 40% energy reductions on specialized smaller models. KPMG's inaugural Global AI Pulse reinforces: 95% of firms have strategies, 8% have established ROI.

The scaling assumption that dominated 2024-2025 is being replaced by an efficiency-and-fit thesis in the data. For product builders: smaller curated-dataset models outperforming massive LLMs on specialized tasks is now a procurement posture, not a research finding. Sovereign-cloud spending signals structural renegotiation of US API dependency at the national level.

Verified across 3 sources: AI Journal (Stanford Index analysis) · KPMG Global AI Pulse · Digital Information World

European AI Chip Startups Raise Nine Figures on Inference-Efficiency Thesis β€” Euclyd Claims 100x Power/Perf vs Nvidia

Euclyd (founded by ex-ASML execs, in talks with four potential customers, claims 100x power efficiency), Optalysys, Fractile, and Arago are all raising nine-figure rounds on the premise that inference workloads have fundamentally different efficiency curves than training β€” and that photonics and data-centric compute can beat GPUs at inference at scale. Geopolitical tailwinds (US export controls, TSMC concentration risk) are accelerating European sovereign-compute capital.

Inference, not training, is where AI economics will be decided over the next 3-5 years β€” long-running agents, always-on copilots, edge deployments. If even a subset of these 100x efficiency claims hold at production scale, the GPU-centric stack assumption breaks for inference workloads specifically. Worth tracking customer announcements over the next two quarters; any named Fortune 500 design wins would be the real signal.

Verified across 1 sources: CNBC

AI Supply Chain & Logistics

Fizyr Vision AI Tackles Logistics Edge Cases β€” Crushed Parcels, Variable Pallets, Domain-Expert-Trained Neural Nets

TU Delft spinout Fizyr deployed vision AI with Pallet Sorting Systems (Heerenveen) for a fully automated sorting tunnel handling crushed parcels, variable pallet dimensions, and adversarial lighting. The methodology: pair domain experts ('Arie the pallet expert') with ML engineers to encode intuitive human judgment into training data β€” prioritizing better data over more data.

Where this week's AGIBOT G2 and Cainiao ZeeBot coverage focused on hardware throughput, Fizyr is the differentiation layer: perception for the long tail of edge cases. Treating domain expertise as a first-class training input rather than a QA step is the pattern that generalizes to any physical-world AI deployment.

Verified across 1 sources: IO+

Spokane & North Idaho

Continuous Composites Expands in Coeur d'Alene Commerce Park β€” Advanced Manufacturing Growth in North Idaho

Coeur d'Alene-based Continuous Composites Inc. β€” which builds carbon fiber parts and the machines that make them β€” relocated to larger office and warehouse space at 3875 N. Schreiber Way in the Coeur d'Alene Commerce Park, with workforce expansion planned.

Continuous Composites' CF3D (continuous fiber 3D printing) technology is one of the more serious composite additive-manufacturing plays nationally, with aerospace and defense customers. A local expansion here is a real signal about North Idaho's ability to retain advanced-manufacturing anchor tenants rather than watch them relocate to Seattle, Boise, or the Bay. Worth watching alongside the broader Inland Northwest industrial-capacity story (Spokane County comp-plan update, Washington data-center power review).

Verified across 1 sources: Spokane Journal of Business

Spokane Transit Authority: May 1 Filing Deadline Forces August-vs-November Call on $30M Sales Tax Renewal

STA deferred its decision to April 29 on whether to place a 0.2% sales tax renewal on the August or November 2026 ballot, with a hard May 1 filing deadline. August placement strengthens the competitive case for $82M in federal Bus Rapid Transit funding; November avoids ballot-crowding with an expected Clean & Healthy Task Force jail/crisis-response tax. Revenue at stake: ~$30M annually.

This decision cascades into BRT federal-match timing and Spokane's transit/development planning under Plan Spokane 2046 β€” the same comprehensive plan update we've been tracking. If STA goes August and the Task Force measure also lands in November, voters see two transportation/public-safety taxes in quick succession β€” historically a recipe for both failing.

Verified across 2 sources: The Center Square · Spokane Public Radio

Washington Awards $5.4M to Seven Spokane County Early Learning Projects β€” Salish Immersion School Gets $1M

Washington awarded $5.4M across seven Spokane County early-learning organizations. Largest grants: $1.69M to Guardian Angels Childcare Center and $1M to the Salish School of Spokane (the state's only full-time Colville-Okanagan Salish immersion program), which will build a new campus targeted for 2028 completion and expand capacity from 18 to 30 students.

Childcare capacity is a leading indicator for workforce participation β€” particularly in lower-income households β€” and the state is currently short roughly 2,000 slots. The Salish component also preserves an irreplaceable language-immersion program that has no equivalent elsewhere in Washington.

Verified across 1 sources: Spokesman-Review

Newport Beach

Costa Mesa Loses Baycrest Caps & Corks After 65 Years β€” Lease Non-Renewal Despite 20% Above-Ask Offer

Baycrest Caps & Corks Bottleshop closed April 1 after 65 years β€” nearly three decades under the same family. A new property owner declined to renew the lease despite the operators offering 20% above asking rent.

The mechanism connects directly to this week's Costa Mesa overlay-zoning story (5,863 housing units withdrawn): both reflect an OC commercial real estate environment where new ownership is overriding established tenants and prior expectations. The pattern is reshaping Newport/Costa Mesa corridors as older retail properties change hands.

Verified across 1 sources: Los Angeles Times (Daily Pilot)

Iran Conflict

Iran Blockade Day 6-7: Hormuz Flips from 'Reopened' to Live Fire on Tankers in 24 Hours; IRGC–Foreign Ministry Split Surfaces

Into Day 6-7 of the blockade: Iran's foreign ministry announced Hormuz 'completely open,' eight tankers transited for the first time in seven weeks β€” then the IRGC publicly contradicted the foreign ministry, laid out strict transit conditions, and Iran fired on at least two tankers. Trump claimed Iran agreed to 'virtually all' US demands; Iran's chief negotiator Ghalibaf denied all seven specific claims. 23 ships turned back since blockade start; a French UNIFIL soldier was killed in southern Lebanon. Ceasefire expires April 22.

The IRGC–foreign-ministry split is the key new development: whatever Araghchi signs may not bind the forces that actually control the strait β€” which means the secondary-sanctions financial strategy Treasury announced on Day 5 may be negotiating with the wrong principal. The 24-hour whiplash is now the pattern. ISW's assessment that Iran will use Hormuz as ongoing leverage rather than genuinely reopen it is the operating assumption to hold going into the April 22 ceasefire expiration.

Verified across 6 sources: Reuters · Institute for the Study of War · Associated Press / Washington Post · Boston Globe · CNBC (central banker reactions) · The National

OSINT & Intelligence

Iran's AI-Generated Lego-Style Propaganda Goes Viral on Western Platforms β€” Asymmetric Influence at Scale

Iran has deployed millions of AI-generated Lego-style videos mocking Trump, Netanyahu, and US military posture across Western social platforms, with surprising cultural sophistication and reach β€” running in parallel with the kinetic blockade conflict.

This is the influence-layer complement to the CSIS autonomous-drone report we covered last week (US-origin AI components as hardware force multipliers). The novel signal for OSINT practitioners: millions of culturally-tuned assets at near-zero marginal cost breaks existing content-moderation and attribution economics in ways that OSINT Navigator and similar tooling isn't yet designed to handle at this production speed.

Verified across 1 sources: RTÉ News


The Big Picture

Agent security is catching up to agent capability Cursor's NomShub RCE, MemoryTrap (earlier this week), and Endor Labs' 7.8% security benchmark all point the same direction: the agentic tools shipping fastest are also expanding the supply-chain attack surface most aggressively. Security is now the gating function, not capability.

Design tooling is getting unbundled into AI subscriptions Claude Design (bundled into Claude.ai), Canva AI 2.0 (agentic workflows), and Google Stitch's March overhaul are all collapsing the design-to-code pipeline into conversational surfaces. Figma's stock reaction (-7-8% across launches) is the market pricing in platform risk.

Cursor's $50B sets a new B2B scaling ceiling Three years from $0 to $2B ARR, 70% Fortune 1,000 penetration, and margin recovery via proprietary Composer models. The playbook β€” own part of the stack to escape API-provider margin compression β€” is now the template for AI-native infra plays.

Hormuz is the new economic leverage point The strait flipped from 'reopened' to live fire on tankers in under 24 hours. IRGC factional disputes with Iran's foreign ministry suggest unpredictable execution regardless of what Trump or Araghchi announce at the podium.

The Inland Northwest is quietly industrializing Continuous Composites expanding in Coeur d'Alene, $5.4M in early-learning grants hitting Spokane County, and Washington's data-center power-demand review (April 27 workshop) all point to infrastructure decisions that will shape the region for a decade.

What to Expect

2026-04-21 Washington UTC written comments due on data-center power demand policy
2026-04-22 US-Iran ceasefire expiration; oil sanctions waiver also expires April 19
2026-04-27 Washington UTC technical workshop on utility handling of large new power demands
2026-04-29 Spokane Transit Authority board decision on August vs. November sales tax renewal ballot (May 1 filing deadline)
2026-05-04 Washington candidate filing week opens for 3rd Legislative District and other races

Every story, researched.

Every story verified across multiple sources before publication.

🔍

Scanned

Across multiple search engines and news databases

611
📖

Read in full

Every article opened, read, and evaluated

131

Published today

Ranked by importance and verified across sources

14

β€” The Anvil

πŸŽ™ Listen as a podcast

Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.

Apple Podcasts
Library tab β†’ β€’β€’β€’ menu β†’ Follow a Show by URL β†’ paste
Overcast
+ button β†’ Add URL β†’ paste
Pocket Casts
Search bar β†’ paste URL
Castro, AntennaPod, Podcast Addict, Castbox, Podverse, Fountain
Look for Add by URL or paste into search

Spotify isn’t supported yet β€” it only lists shows from its own directory. Let us know if you need it there.