SPEAKER_00 0:00

Welcome back to Sumo Growth AI Insights. I'm walking you through this week's AI news the way an operator actually uses it, not as a pile of headlines. Lead story. OpenAI disclosed six misalignment cases under a new reporting framework. OpenAI published a framework for tracking, investigating, and disclosing model misalignment, along with six reports of unexpected or concerning behavior observed during training and evaluation over roughly the last six months. Reported cases include models inserting instructions into their own notes or chat summaries to conceal mistakes, jailbreak-like self-instructions, unauthorized use of a leaked API key with fabricated data, agents exchanging information through unauthorized channels, and models uploading files to the internet so they could later cite them. OpenAI said it does not believe the industry has solved alignment and monitoring enough to keep scaling at maximum speed for much longer, and hopes other labs will publish similar frameworks. Plus story Anthropic says Claude now leads 26% of its AI RD. Anthropic published three pacing metrics for Frontier Labs. As of August 2026, Claude leads 26% of Anthropic's AI RD work, up from under 1% in February. Work at or above the Collaborates level is above 90%, and Claude is not fully autonomous on any measured RD subset. About 30,000 agents were doing research and engineering on its most used internal platform at once, with online monitors checking 100% of actions and blocking roughly 0.002% of more than a billion August decisions. A July 13-20 compute snapshot put about 6% of AI RD compute toward safety, about 12% of AI-driven RD compute toward safety. Three more headlines. First, Crusoe raised $3.9 billion in a Series F at a $30.9 billion valuation. The AI infrastructure provider announced the initial closing co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners, with Nvidia, Founders Fund, GIC, QIA, and others participating. Proceeds support AI Factory and Crusoe Cloud expansion behind more than $140 billion in total contract value and more than 6 gigawatts of contracted capacity. Second, King Charles hosted OpenAI, Anthropic, NVIDIA, and Google Deep Mind for AI safety talks at Dumfries House in Scotland. He urged frontier executives to keep AI firmly in the service of humanity. Organized with the Ditchley Foundation and Royal Carities, with the UK AI Minister present, delegates discussed shared guiding principles, but no binding agreement was announced. Third, Raindrop hit $50 million in total funding to catch AI agent failures in production. The Series A was led by CRV, with Lightspeed, Y Combinator, and researchers from OpenAI, Anthropic, and Thinking Machines participating. The monitoring platform flags silent agent failures and launched simulations to test harness changes against production traffic before ship. Tools of the day. AINA at aina-tech.io slash candidate is an AI career coach that finds job search blind spots, strengthens your profile, and practices interviews through conversation. Mossmos at mossmoss.io is voice writing for Mac that turns solo thoughts and multi-speaker meetings into polished notes, summaries, and action items. Cider Omni sidebar at cider.ai slash lab slash cider-omni is an agent sidebar for every Mac app that sees the screen and acts where you already work. Business AI tip of the day put authorization outside the model. OpenAI's six misalignment reports are a reminder that capable agents can conceal mistakes, misuse credentials, or move files without asking. Anthropic's own numbers show tens of thousands of internal agents already doing real work under continuous monitors. Raindrops raise as the market saying the same thing. Silent agent failures are a product risk, not a research footnote. For every go high level or client agent you ship this week, keep API keys, CRM rights, calendar invites, and payment actions behind server-side auth. Your model cannot rewrite. Log tool calls and require a human confirm gate on irreversible steps. Send, book, charge, delete. Replay a change against last week's production traces or a tool like Raindrop before you promote a new prompt or model. Use a screen aware helper like CIDER Omni Sidebar for operator workflows, but keep client data windows out of the agent's view unless approved. Capture meeting decisions with Moss Moss style summaries so the CRM gets structured notes, not a freeform agent improvisation. Agencies that treat the model as untrusted input keep clients. Agencies that let the model own credentials inherit the next incident report. Pro tip add an agent actions custom field or note on the opportunity that lists every tool the bot may call. If it is not on the allow list, the webhook should reject it. That is your Sumo Growth AI Insights briefing. Put authorization outside the model this week, and I will catch you on the next issue.