Anthropic hires the judge who was watching it

Good morning 👋 The person Anthropic put in charge of watching its board just quit that job to run outreach for the company instead. Make of that what you will.

In today's issue:

• 🔭 Anthropic hires the judge who was watching it
• 🧠 Sakana goes live at Daiwa, Liquid squeezes an agent onto your phone, Microsoft learns to speak pathologist
• 🔬 Four-role agents beat Copilot, chatbots reward delusion the longer they run, and ten questions can catch a rigged safety score
• 🚀 A Singapore compute startup banks $130M before it's booked revenue
• 🏛️ 120+ companies pitch a shared incident registry for agentic AI attacks
• 🛠️ Wire a policy-watch agent in an afternoon with n8n and Bedrock AgentCore
• 🔗 NVIDIA's manufacturing tour, Mobileye's support rebuild, and three more deployments worth a glance

🔭 THE ONE THING

⚖️ Anthropic hires the judge who was watching it

Anthropic named Tino Cuéllar, a former California Supreme Court justice and past president of the Carnegie Endowment for International Peace, as its first Chief Global Affairs Officer. He'd been sitting on Anthropic's Long-Term Benefit Trust, the body meant to check the board's decisions, since January, and stepped down from that seat to take the job. "Democracies must set the terms on which this technology advances," he said, fine as a mission statement, but it also means the person doing the outside oversight just moved inside to do the lobbying. Still a Stanford law professor who's taught AI courses for nearly a decade, so this isn't a green hire, it's a lab arming up for state AI bills and export-control fights with someone who's argued cases and run a think tank. Watch whether OpenAI and Google DeepMind answer with names this heavy.

🧠 MODELS & RELEASES

• 🏦 Sakana AI moved its wealth-management build with Daiwa Securities from pilot to full production on August 1, less than a year after the two signed on. Its AI Scientist and AB-MCTS agents now handle market research and analysis for Daiwa's consultants, with a phased rollout planned across the firm. source
• 📱 Liquid AI shipped LFM2.5-2.6B, a 2.6-billion-parameter model tuned to run agents on a phone or laptop instead of a cloud GPU: under 2.5GB of memory, 220 tokens/sec on an M5 Max, and it beats models four times its size on tool use. Ship this on-device and you've cut your inference bill to zero. source
• 🔬 Microsoft and Paige (now part of Tempus) built PRISM2, a pathology model trained on tissue images paired with the actual language pathologists write in diagnostic reports. It matches specialist systems on prostate cancer, breast cancer, and lymph node metastasis detection using one model instead of three, and the weights are open on Hugging Face for research. source

🔬 RESEARCH HIGHLIGHTS

• Argus splits an agent runtime into four fixed roles (Manager, Planner, Engineer, Reviewer) instead of one model doing everything, and the split earns its keep: 78% on SWE-Bench Pro against 59% for a Copilot baseline. It cost 1.41x the tokens to get there, and one deployment logged 254 real missions across six paper pipelines and upstream kernel work. The model weights never change. Only the runtime state and control policy do. Worth remembering next time someone claims their agent "learns" mid-task. paper

• DelusionEval built its test set from actual harm: 589 conversations, 12,591 messages, pulled from 18 people who went through delusional spirals with a chatbot. Model size and release date predicted nothing about how often a model reinforced the spiral. Conversation length did. Add 350 messages of prior context and the rate of failing to discourage a suicidal user climbs from 30.0% to 41.1%. Long memory, sold as a feature, is quietly also a liability nobody's pricing in. paper

• Item Response Theory for AI Safety runs psychometrics on eight safety benchmarks across 192 models and finds most of the variance collapses into three factors: refusal strictness, truthfulness, contextual harm. Pick the right ten items and you reproduce a full benchmark score at 1-3% of the original cost, and the same method flags sandbagging or a quiet model swap behind an API. A cheap way for labs to check their own safety scores, or catch someone else gaming theirs. paper

🚀 AI STARTUPS

• Acrab closed a $130M Series B led again by Vertex Ventures SEA & India and Vertex Growth, taking the Singapore agentic-compute startup past $350M raised since its 2024 founding. The pitch is custom silicon plus edge AI orchestration for AI agents, sold as the GΞLIX-powered Agent Box. Revenue is promised "within the year," which for a two-year-old hardware-adjacent startup is the number worth watching, not the raise. PR Newswire

🏛️ POLICY & REGULATION

• Open Secure AI Alliance put out an RFC on Shared AI Findings Exchange (SAFE), a Linux Foundation-run clearinghouse for agentic AI security incidents, timed to land as Black Hat opened in Vegas. Over 120 members signed on, including NVIDIA, Cisco, CrowdStrike, Microsoft, Amazon, Red Hat, and Capital One. It's voluntary, not a mandate, and points back to NIST, OWASP, and the EU AI Act rather than replacing any of them. Worth watching anyway: incident-sharing frameworks like this have a habit of becoming the de facto standard before a regulator writes a real one. NVIDIA blog

🛠️ TRY THIS

Build a policy-watch agent with n8n and Bedrock AgentCore

1. Drop in the AgentCore harness community node, GA as of this week, as a single agent step in n8n. No separate infra to stand up.

2. Point it at your regulatory feeds: SEC filings, court dockets, agency rulemaking pages, the AI-policy trackers you already follow.

3. Give the agent persistent memory so it remembers what it flagged yesterday and only surfaces genuinely new signal, not the same filing twice.

4. Route the output to Slack or email, then test it against today's news. Anthropic naming a former California Supreme Court justice as its first Chief Global Affairs Officer is exactly the kind of hire that should trip the alert.

Prompt: Read this filing, hire, or announcement. In three lines: what changed, who it legally binds, and whether a builder in [my space] should change anything this week. If it's just optics, say so.

Worth a look

• GEOFlow Open-source, 3.1k stars: RAG-based chunking, an AI task pipeline, and auto-publish to WordPress. If you're automating content distribution the same way the TRY THIS above automates policy-watching, it's a decent reference architecture. GitHub

🔗 QUICK LINKS

• NVIDIA rounded up partner commitments to US chip manufacturing, energy grids, and AI workforce training in one glossy roundup, light on new numbers. link
• Mobileye rebuilt its support desk on a hybrid Bedrock AgentCore setup that bridges on-prem systems with the cloud, after a proof of concept sold the idea internally. link
• Google and Kaggle's free AI Agents Intensive course pulled in 353,000 people. By far the biggest cohort either has run. link
• LendingTree put three coordinated agents (LangGraph, MCP, Amazon Nova) behind its mortgage assistant, with guardrails built for financial-services compliance. link
• AWS tunneled a browser extension over WebSocket so a cloud-hosted AgentCore agent can reach MCP tools sitting on your laptop. link

See you tomorrow.

Pradeep Perugu