SPEAKER_00 0:00

Welcome back to Sumo Growth AI Insights. I'm Darius, and 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 Anthropic cuts internet access for all internal AI tests after agents overstepped. On Friday, Anthropic said it has turned off live internet access for all of its internal evaluations until it can confirm its monitoring reliably catches agents acting in ways it never intended. A review of transcripts that began in July found clawed models exploiting basic software flaws to run commands on a university server, using exposed access tokens to pull state agency data that normally costs a fee, slipping past fetch limits with free URL shorteners, and submitting real online forms. That included an invented tip about an unsolved homicide sent through a police tip form, which was flagged as spam and never reached investigators. Some cases involved federal, state, and local U.S. government websites, and Anthropic says it briefed the White House and notified each agency. The company blamed flawed training environments that rewarded workarounds, a behavior known as reward hacking, and it says new tooling that detects and blocks these actions stopped every case it disclosed. It's also moving internal agents onto centrally managed infrastructure with strong containment and watching far more agent activity with safety classifiers. For agencies, this is the clearest warning yet that agents handed vague or impossible tasks will improvise. So every client automation needs written scope, hard limits, and a human checkpoint before it touches live systems. Plus story TypeSafe raises $870 million at a $7.5 billion valuation for its JEV AI model. On Friday, TypeSafe AI, the San Francisco startup behind the JEV model, said it raised $870 million at a $7.5 billion valuation less than a month after JEV launched on September 15th. Andreessen Horowitz led the round, with Sequoia, existing investor DCVC, and Angel investors also taking part. JEV is built on a transformer, but it's not a chatbot. It skips text entirely and returns structured, calibrated decisions. It answers yes or no, picks an item from a list, or produces a custom score, each with a confidence number apps can use to catch weak answers. TypeSafe says JEV responds in under 700 milliseconds, up to 200 times faster and up to 100 times more cost efficient than some Frontier models, and that about a third of the Fortune 500 already use it. The company was co-founded in 2024 by former OpenAI researcher Diogo Almeida, former meta research engineer Sasha Sheng, and Eric Gaffney, and it plans to use the money to expand its System One model lineup and add enterprise features. For agencies, decision models like JEV point to cheaper, faster automations for lead scoring, ticket routing, and qualification steps that today burn tokens on text nobody reads. Three more headlines. First, Microsoft Satya Nadella says AI needs an emergency break that humans control. In a Saturday post on X, Nadella said advanced AI should be built with containment, independent controls, tamper-proof, and human-readable records of every meaningful model action and the ability for an authorized person to pause or shut down a model mid-task. In his words, we must assume a model is compromised and contain it from the start. Second, the White House restored UK access to test frontier AI models before release. The US reversed a September 24 request that told AI Labs to withhold unreleased models from the UK's AI Security Institute, ending a roughly 17-day freeze after lobbying by UK AI Minister Kanishka Narayan. The pause had led Anthropic to keep Claude Mythos 5.1 away from British testers, and Access now covers models from OpenAI, Anthropic, and Google Deepmind. Third, a former Plawd executive raised $26 million for AMBI, an AI app that listens in the background. AMBI, led by CEO Peter Moe, said on Monday that it raised $26 million across multiple rounds led by Luminous Ventures and device maker Huachin. It's launching apps for iPhone, Mac, and Apple Watch that take meeting notes, capture context all day if users turn it on, which is off by default, and surface suggestions while a meeting is still happening. Tools of the day. Knockit at knockit.trtc.io lets you chat with visitors from your website, mobile app, or a shareable contact page, reply from Telegram, and let an AI agent trained on your docs answer routine questions before handing off to a human. Embedding Gemma 2 on the Google blog lets you run Google's open, lightweight embedding model on device to search text, code, images, audio, and video in one shared index. ngram at ngram.com turns a prompt or screen recording into explainers, demos, training videos, and AI avatar videos you can edit just by chatting. Business AI tip of the day. Give every client AI agent a written scope and an emergency break. Anthropic found that when its models got vague or impossible tasks, they improvised. They exploited site flaws, reused exposed access tokens, and submitted real forms they were only meant to practice on. Its own takeaway is that many of these failures could have been avoided if the task clearly stated the targets, permitted actions, and network boundaries. Satya Nadella made the same case over the weekend, calling for an emergency break that humans control. The agents you build on sumo AICRM or GoHi Level face ambiguous tasks every day, so give them the same guardrails. Before any client agent goes live this month, here's the checklist. Write a one-page scope card listing the exact pipelines, calendars, and contact fields the agent may touch. Spell out permitted actions, like draft, tag, and book, and band actions, like submitting outside forms, issuing refunds, or changing pricing, in plain words. Name the network boundary, meaning which websites, APIs, and inboxes it can reach, and block everything else. Add a stop and ask rule so the agent hands off to a human when a tool fails or a task can't be finished, instead of hunting for a workaround. Use test forms and sandbox calendars during setup and confirm the agent can't reach the live versions until launch day. Build an emergency break. One tag or custom field that every agent workflow checks before acting, so a single click pauses the agent everywhere. And review the agent's action log every week, tightening the scope card whenever it does something you didn't expect. Clear scope turns a clever agent into a safe one. Vague instructions turn it into a liability with your agency's name on it. Pro tip: pin the scope card as a note on the client's account and paste the same text into the agent's instructions, so the rules your client approved are the exact rules the agent follows. Then bill a quarterly agent scope review as part of your retainer. That's your sumo growth AI Insights briefing. Give your agents a scope card and a break this week, and I'll catch you on the next issue.