Anthropic Drops a Claude Product for Mom-and-Pop Shops

Anthropic launched Claude for Small Business, a new bundle inside Claude Cowork aimed at the 36 million US small businesses that account for 44% of GDP. Toggling it on unlocks bookkeeping, business insights, and ad campaign generation, plus integrations with QuickBooks, Canva, Docusign, HubSpot, and PayPal. The company is running a 10-city promotional tour with free training workshops for 100 small business leaders per stop, chasing the customer class OpenAI has been quietly grabbing since 2023.

Microsoft unveiled MDASH, an agentic security system that orchestrates more than 100 specialised AI agents across frontier and distilled models to find, debate, and prove exploitable bugs in Windows. The system flagged 16 previously unknown vulnerabilities, four rated critical RCEs in TCP/IP and IKEEXT, all included in this Patch Tuesday. MDASH opens to enterprise customers in private preview in June.

Akamai analyst Tomer Peled tracked down vulnerabilities in MCP servers for three popular database projects that could let attackers run arbitrary SQL on Apache Doris, exfiltrate metadata from Alibaba RDS, and take over internet-exposed Pinot instances. Apache shipped a CVE and patch for Doris MCP, and there's an open ticket on Pinot. Alibaba refused to patch RDS MCP, saying the missing validation isn't a bug.

Anthropic CFO Krishna Rao said Claude now writes 90% of the company's code and handles most of its finance work, collapsing hours of office work into minutes. Execution roles have become supervision of AI, requiring higher talent density across every function. The work is changing from doing to overseeing, Rao told Patrick O'Shaughnessy on Invest Like the Best.

Cisco beat Wall Street's Q3 expectations on surging AI orders, sending shares up 17% in extended trading. The company will cut roughly 4,000 jobs as part of an AI-driven restructuring. New networking switches and routers shipped during the quarter, and the segment that has trailed data-centre rivals through the AI boom is finally catching its share.

Anthropic released Claude for the legal industry on May 12, the first half of a two-product week that also brought Claude for Small Business on Wednesday. The Legal launch ships 20+ MCP connectors linking Claude to contract lifecycle systems, research platforms, document management, e-discovery, and data rooms, plus 12 plugins tailored to specific practice areas. Free Law Project and the Justice Technology Association are partners.

Notion launched a developer platform that extends its custom AI agents, lets them connect with external agents, and runs multi-step workflows pulling data from any database. Customers have built over a million Notion agents since February. New Workers run code in a sandboxed cloud, webhooks trigger from other apps, and Notion positions itself as the hub where humans and agents collaborate.

Cline shipped @cline/sdk, the agent runtime that powers its VS Code and JetBrains extensions, available as a portable npm install for any team building coding agents. The team rebuilt the core agent loop from scratch as a standalone SDK and migrated their CLI and Kanban on top of it. Provider-switching becomes a config change instead of a fork, with shared, llms, agents, and core layers each replaceable in isolation.

Anduril closed a $5 billion Series H led by Thrive Capital and Andreessen Horowitz, doubling its valuation to $61 billion and shattering recent defence tech funding records. The defence giant is also standing up an autonomous naval shipyard in Seattle, expanding from drones into surface warships. The round signals where dual-use ML talent and infrastructure budgets are clustering.

Spotify open-sourced a Claude Code plugin that turns natural-language requests into Spotify Ads API v3 calls, available on Anthropic's Claude Plugins marketplace and GitHub. The Ads API has 30+ resource types and nested targeting structures, exactly the kind of high-cognitive-distance interface agents are good at smoothing. The implementation is Markdown skill files, one bash script, and two small Python helpers, deliberately simple so teams can fork it.

When Databricks launched real-time model serving, a single customer could blow past what its Envoy plus Ratelimit plus single-Redis architecture was designed for, with P99 climbing to 10-20ms across two network hops. Adding machines and caches stopped scaling. The redesign answers the client as if the decision is already made and reconciles state lazily, giving up strict accuracy on the hot path. That trade is the reflex more teams should reach for.

Google's SIGMOD 2026 paper shows how to accelerate AI-powered SQL functions with proxy models, ultra-lightweight task-specific models that replace most LLM calls during query execution. They consume cached embeddings rather than recomputing semantics per query, run on CPU, and ship in BigQuery and AlloyDB optimized mode. F1 holds at 90 to 116% of pure-LLM quality across ten benchmarks, with the technique breaking on multi-step reasoning.

Alibaba's Qwen team released Qwen-Image-2.0, an omni-capable image foundation model that couples Qwen3-VL as condition encoder with a Multimodal Diffusion Transformer for joint condition-target modelling. The model accepts instructions up to 1K tokens and targets text-rich content like slides, posters, infographics, and comics. Significant improvements landed in multilingual typography and photorealistic lighting, and human evaluations show it outperforms previous Qwen-Image versions on complex prompts.

Fully automated codebases scare builders because they treat agent output like junior developer code, says Skip Labs. The fix isn't more manual review, it's building the infrastructure that makes agent code trustworthy: formal specification layers, non-determinism testing, AI verification pipelines, and production monitoring with rollback. Once agent output is treated like compiler output rather than human first drafts, the volume becomes the feature, not the bug.

Coding reached AI ubiquity in 2.5 years because verifiability was easy, tests pass or fail. The bottleneck has shifted from model capability to industry-specific friction. Finance has the same hunger for automation but lacks the cheap, instant feedback loop coding got from compilers and unit tests. There's a brief arbitrage window for builders who can build the verification layer for industries that don't yet have one.

Steve Bannon and Bernie Sanders agree on almost nothing, but both think AI is a disaster for the working class. Maine just passed the first statewide data centre moratorium. Last month someone fired 13 rounds at an Indianapolis councilman's house with a 'NO DATA CENTERS' note, and a man threw a Molotov at Sam Altman's home. Researchers call this the structural conditions historically associated with political violence.

If you strip away the UI and expose the database, what are you actually left with. a16z argues defensibility is moving in agentic software: UI stickiness dies when humans aren't the primary user, and value relocates down to the data layer and up to execution. Salesforce going headless is the tell. Builders need to rethink where moats sit when agents, not humans, drive the interface.

MinerU is an open-source document extractor that converts PDFs, slides, and Office files into structured Markdown and JSON shaped for agent and RAG workflows. It preserves tables, formulas, and reading order, the parts where commercial OCR usually breaks. The repo is trending on GitHub with 62.9k stars and active commits from OpenDataLab. If your RAG pipeline is duct-taping Unstructured.io and pdfplumber, this is worth a Saturday afternoon to swap in.

Supertone open-sourced its TTS engine and picked up over a thousand GitHub stars in a single day. The model runs natively via ONNX with no cloud round-trip, no per-character billing, and no PII leaving the device. The reference implementation is Swift-first for iOS, but ONNX runs anywhere a runtime loads, including mobile, desktop, and edge. The first credible local alternative to ElevenLabs in months.

Microsoft Research's concept-driven visualization tool went trending again at 15.5k GitHub stars. You describe what you want to see, like compare quarterly revenue against headcount faceted by region, and the AI handles data reshaping, encoding, and chart generation. Iterative refinement lets you nudge it like Claude in Artifacts. The closest commercial equivalents are Julius and ChatGPT data analyst, but Data Formulator is fully open source and runs against your own LLM keys.