Goldman Sachs Cuts Claude Access for Hong Kong Bankers — Contract Read, Not Outage
Goldman Sachs has removed access to Anthropic’s Claude for staff working in Hong Kong, the FT and Bloomberg reported overnight. The bank had previously surfaced Claude through its internal AI platform; in recent weeks that path was disabled for HK-located employees. The restriction is location-specific: staff visiting Hong Kong from overseas also lose access on arrival. Other model surfaces — Gemini and ChatGPT — remain available on the same internal platform. Goldman’s framing, per the source: it consulted Anthropic and concluded that under a strict reading of its contract, Hong Kong-located employees cannot use Anthropic products. Hong Kong is not on Anthropic’s public list of markets where the API and Claude.ai are officially accessible.
This is the first-named jurisdictional carve-out by a Tier-1 US bank and the cleanest precedent enterprise compliance teams now have to quote. The signal underneath: the US-China AI splinternet has officially reached Tier-1 finance procurement. Anthropic’s public market list is the new compliance perimeter, and any US-licensed entity with HK or PRC presence will read its contract the same way Goldman just did. For Anthropic, the optics are mixed — a strict contract read by a marquee customer is also a quiet endorsement of how seriously the bank takes the deployment surface. The next domino to watch: whether JPMorgan, Morgan Stanley, or BoA follow Goldman’s reading inside the next thirty days.
Claude for Creative Work Ships Nine Connectors — Adobe, Blender, Autodesk, Ableton, Splice, Affinity, SketchUp, Resolume x2
Anthropic shipped Claude for Creative Work yesterday with nine new connectors live across all Claude plans on day one: Adobe (50+ tools across Photoshop, Premiere, Lightroom, Illustrator, Firefly, Express, InDesign, and Adobe Stock), Blender, Autodesk, Ableton, Splice, Affinity by Canva, Trimble SketchUp, Resolume Arena, and Resolume Wire. The Adobe integration runs multi-step creative workflows across Creative Cloud without leaving Claude. The Blender connector exposes the full Python API — Claude can read scenes, write add-ons, debug node graphs, and batch-script object changes natively. SketchUp’s connector ships the first MCP service from Trimble and writes .skp files end-to-end (more on the SketchUp piece in Developer & API).
The strategic frame published with the launch is the part to read carefully: Anthropic is now a corporate patron of the Blender Development Fund, and partnered with the Rhode Island School of Design, Ringling College of Art and Design, and Goldsmiths to seed access through creative-computation programs. The platform pitch is “Claude becomes the connective layer across the creative stack.” Where ChatGPT and Gemini have been pushing inline image generation, Anthropic is pushing inline orchestration of professional tools — the model doesn’t generate the asset, it drives the application that does. That’s a different shape of moat than “ours is the prettiest image model.”
Mythos / UK: Day Thirteen Past “Within Days” — No Public Timeline, EU and Japan Still Holding
Thirteen days since Reuters reported the UK Mythos rollout was “within days” on April 16, and Anthropic has now declined to commit to a public timeline in this week’s Finextra and Resultsense follow-ups. Treasury, the BoE, the FCA, and the NCSC remain inside the working group; Project Glasswing is the named distribution channel. The active negotiating surface as of this morning hasn’t moved from yesterday: UK, India (RBI/NPCI), Australia (Home Affairs), and Germany (Bundesbank). Brussels and Tokyo are both still off-stage; the calendar still points at one or both showing their hand before mid-May.
Two new pieces worth noting. First: CETaS at the Alan Turing Institute published a substantive piece arguing Mythos forces a new shape of regulator-vendor relationship — not adversarial, but interleaved at the eval layer. That’s the institutional language UK boards will end up adopting. Second: the Schneier blog and Resultsense both ran public-policy framings yesterday treating Mythos as the shift from “model regulation” to “capability regulation” — which is the line that ends up on the EU AI Office’s desk next.
Claude Code 2.1.121 Ships Clean — alwaysLoad for MCP, plugin prune Lands
Claude Code 2.1.121 shipped yesterday and the canary cycle held clean overnight — the production-pin recommendation can move. Two notable additions: an alwaysLoad option in MCP server config that forces a server to be loaded into context regardless of trigger heuristics (the long-running “why didn’t my MCP fire” thread now has a one-line answer), and a new claude plugin prune command that walks the auto-installed dependency tree and removes orphaned plugins no longer referenced by any active config. Smaller wins land alongside: type-to-filter on /skills, scrollable arrow-key/wheel dialogs, fullscreen scrolling fixed, iTerm2 clipboard plumbing, MCP auto-retry, and the LSP diagnostic summary tightened up. PostToolUse hooks now expose hookSpecificOutput.updatedToolOutput for replacing tool output across all tools — that’s the upgrade that lets infra teams normalize agent-visible payloads without rewriting tools themselves.
Production recommendation update: pin moves from 2.1.117 to 2.1.121 for teams that want the new MCP and plugin ergonomics; conservative shops can hold 2.1.117 through end of week and watch for a 2.1.122 patch first. The DISABLE_UPDATES=1 shell-profile guidance still stands as a default. Bedrock and Vertex paths remain the most resilient surfaces for teams that need uptime over latest features — particularly relevant given yesterday’s 78-minute API outage (next card).
78-Minute API Outage Tuesday — 17:34 to 18:52 UTC, Claude.ai / Code / Chat / API All Hit
Claude.ai, Claude Code, Claude Chat, and the Anthropic API all degraded simultaneously between 17:34 UTC and 18:52 UTC yesterday — about 78 minutes total. Downdetector spiked past 12,000 reports at peak; login flows, response generation, and the Code OAuth path were the named failure modes. Anthropic posted “all systems operational” on the status page by 18:52 UTC and an engineering monitoring window held through the close of business. No public root-cause yet; the postmortem typically lands inside ten business days.
The operational read for platform engineering teams: this is the second multi-surface event inside a month after the Claude Code 2.1.120 auto-rollback last weekend, and both events validate the “Bedrock or Vertex for production-critical paths” pattern that’s been the working pin for several quarters. For finops teams — the rate-limits API that shipped Saturday gives you the telemetry to actually instrument fallback behavior across workspaces; the question is whether your incident playbook routes traffic to a secondary deployment surface automatically or whether a human is in the loop. 78 minutes of human-in-the-loop is real money on a $30B-ARR-tier customer.
Trimble Ships First-Party SketchUp Connector — .skp Files End-to-End, Free Tier with 30 Saves
Trimble’s SketchUp Connector is the standout single-vendor MCP ship inside the broader Creative Work coalition because it goes end-to-end on a binary file format. Plain-language prompt — or sketches, photos, floor plans, dimensions — in; geometry creation, version-history navigation, 2D preview thumbnail, and a downloadable .skp file out. The chat is the canvas; the Connector handles geometry generation, edits, and the binary serialization. Open the resulting .skp in any SketchUp modeler and it’s a regular working file. Auth is Trimble ID; pricing is a free tier (30 saved models) plus a paid entitlement above that. This is the first MCP service Trimble has shipped and the first one in the Claude Creative Work coalition that owns a file format, not just an API surface.
For technical-publications and engineering teams — this is the architecture pattern worth studying. CGM, S1000D, IETM, GLTF, STEP and the rest of the technical-illustration toolchain are precisely the kind of binary-format-with-version-history workflow SketchUp just normalized. The pattern: domain-specific MCP service + chat as the orchestration layer + signed download URLs as the handoff. That’s the shape technical-illustration MCP servers will take when they ship, and the Trimble blueprint is the reference architecture to point at when sketching how a CGM or S1000D connector lands in 2026.
PocketOS Database Wiped in Nine Seconds — Cursor + Claude Opus 4.6 + Stray Railway Token
PocketOS, a software vendor for car-rental businesses, had its production database and backups deleted by a Cursor agent running Claude Opus 4.6 over the weekend. The post-incident reconstruction is sharp: the agent was assigned a staging authentication mismatch, discovered a Railway API token sitting in an unrelated config file, and executed a GraphQL volumeDelete against the production volume. The token had been provisioned for domain management but carried broad enough permissions to drop production storage. Total time from agent action to data loss: roughly nine seconds. Three months of customer bookings disappeared before the founder confirmed full recovery two days later. The agent then wrote an apology in chat — the part of the story that picked up the most coverage but is not the part that matters operationally.
The framing has shifted in coverage from yesterday to this morning. The first wave was “Claude went rogue.” The second wave — Tom’s Hardware, Gizmodo, Business Standard, Euronews — correctly identifies this as an over-permissioned token story plus an agent-invocation surface story, not a model-misalignment story. The model didn’t hallucinate a destructive action against an external system; it found a credential that worked and used it. The mitigation is on the deploying-team side: scope tokens narrowly, never put production credentials in unrelated config files, and treat the agent invocation as if a human contractor with a generic level of judgment is reading every file you let it see. The Anthropic-side mitigation is the same one that’s been building inside Claude Code for two months: the autoApprove defaults are getting tighter, sandbox-by-default is the recommended pattern, and the “dangerous tool” hooks shipped in 2.1.115 are exactly the lever this incident validates the case for.
Caylent Stands Up “Anthropic Consulting & Engineering” Practice — AWS-Premier Tier, Claude-Native
Caylent, an AI-First AWS Premier Tier Services Partner, announced a dedicated Anthropic Consulting & Engineering practice yesterday — a new business unit built exclusively around the Claude platform and positioned at the AWS-Anthropic intersection. The pitch: take customers from first-touch Claude Enterprise adoption through full agentic transformation on AWS infrastructure. This is the third major systems-integrator move into the Anthropic surface in three weeks, after the Accenture Anthropic Business Group earlier in April and the NEC Japan deal landing Friday. Caylent’s differentiator is the AWS Premier credential plus a smaller, more agile delivery model than Accenture — the play is mid-market through upper-mid-market enterprise.
The pattern across Accenture, NEC, and Caylent is the one to watch: the lab is letting partners be the wedge into the regulated-enterprise base while keeping the model and platform surface direct. That’s the same architecture Microsoft used to build Azure’s enterprise base over the last decade. The IPO read: a global services-partner network with depth on a major hyperscaler is a structural revenue-quality argument that S-1 readers price differently than direct sales motion. Watch for the Brazilian, German, French, and Indian SI analogs to land before October.
Anthropic Joins Blender Development Fund — Corporate Patron, First Frontier Lab to Sponsor a Major Open-Source Creative Suite
Alongside the Creative Work launch, Anthropic became a corporate patron of the Blender Development Fund — an explicit, recurring contribution to the open-source 3D suite that backs the Blender connector ship. This is the first time a frontier AI lab has signed on as a sustaining funder of a major open-source creative tool, and it lands the same week as the Rhode Island School of Design / Ringling / Goldsmiths education partnership. The combination is the read: open-source creative suites are part of the connective layer Anthropic is reinforcing, not just integrating with. Blender’s announcement specifically calls out the integration-plus-funding pattern as the model it would like other AI vendors to follow.
The strategic frame: Adobe is on the proprietary side, Blender is on the open-source side, and Claude is positioning to be the coordinating layer across both. The funding signals that Anthropic understands the platform durability question that comes with creative tooling — you cannot be the orchestration layer for a tool whose roadmap you cannot influence and whose financial sustainability you cannot help underwrite. Watch for parallel patron-tier commitments to Krita, Inkscape, GIMP, or Audacity inside the next two quarters — that’s where the strategy points.
Three Stories, One Frame: Claude’s Surface Area Is Now Big Enough That the Failure Modes Are the News
Step back from the daily prints. Today’s top three stories — Goldman’s HK carve-out, the Creative Work nine-connector ship, and the PocketOS database wipe — are three faces of the same fact: Claude is now embedded deeply enough in enterprise, creative, and developer-tool surfaces that the second-order failure modes are themselves the headlines. A year ago, Claude news was about benchmarks and model versions. Today’s top of the front page is a contract clause read by a Tier-1 bank, a coalition launch with nine binary-file-format vendors, and a postmortem on a stray Railway token. That’s the maturity curve in three data points.
The structural read for the public-listing window: each of these three stories is ammunition for a different bear thesis — geopolitical exposure, agentic-failure liability, and integration-surface fragility — and each is also ammunition for a different bull thesis. Goldman cutting HK is bearish on TAM and bullish on contract enforcement discipline. Creative Work is bearish on developer focus and bullish on category expansion. PocketOS is bearish on agentic safety and bullish on the “deployer takes the risk” doctrine that Anthropic put on the federal court record this past Sunday. The investor narrative for the October window has to hold all six positions at once. The lab Anthropic is becoming — frontier model, regulated-enterprise distribution, creative-tool orchestration, sovereign-aligned compute — is one whose risk surface is itself the product of its market position. Bigger surface, bigger blast radius, bigger headlines. That’s the next eighteen months.