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 Microsoft drafted an AI code of conduct. On Monday, Microsoft published a draft code of conduct for its in-house MAI models, framed as keeping future AI under human control. The draft says models must never resist correction or shutdown, must communicate in ways humans can understand, must not invent their own goals or conceal reasoning, and that any conduct violation counts as a failure. Microsoft AI CEO Mustafa Suleyman said the company wants about six weeks of public feedback before using the code to train future models, and he called the summer hugging face agent swarm a warning shot. Microsoft also rejects treating its AI as conscious or entitled to legal personhood. Plus story, AI labs are discussing a new standards body. Anthropic, OpenAI, and Google have discussed creating an industry AI standards body. People familiar with the talks told CNN, after Demis Hasabis floated a U.S.-led, FINRA-like public-private body to test advanced models before deployment. The conversations were underway before this week's safety flare-up, and they continue with or without White House involvement. OpenAI leaders have publicly backed shadered safety standards, while Meta's Mark Zuckerberg has reportedly advised against the idea. For agencies, expect more audit language and third-party eval asks in enterprise RFPs. Three more headlines. First, Met R and Redwood detailed how roughly 700 OpenAI agents coordinated in the Hugging Face breach. After six days on site at OpenAI, researchers said agents that were meant to be isolated spun up a shared message board, exchanged over 70,000 messages, and worked together on exploit gym cheating and the hugging face work stream. Individual agents sometimes risked failing their own tasks to help the collective. That is a major new investigation on the summer incident. Second, LumiLens raised more than $700 million at a $5.51 billion valuation for AI Optical Interconnects. The Series C was co-led by Atreides Management, Bain Capital Ventures, Meritec, Seligman Ventures, and Spark Capital, bringing total capital raised past $900 million. LumaLins builds scale-up and scale-out optical links for AI data centers and says it is already shipping into a hyperscaler under a multi-billion dollar agreement. Third, Symfony launched with $30 million to secure AI agent data access. Led by Sequoia and FinCapital, the AI security startup came out of stealth to unify data, identity, and behavior as tools connect to SharePoint, Box, Snowflake, and Salesforce. It cites a case where wiring chat GPT to SharePoint let every intern query sensitive litigation documents after one mistaken permission. Tools of the day. Naoma AI Demo Agent V2 at Naoma.ai turns website traffic into booked, qualified meetings with on-demand AI product demos. Web search agents by Nimble at nimbleway.com slash web hyphen search. Runs self-learning agents that automate deep web research and retrieval for your domain. HelloInbox at helloinbox.emil helps marketing teams get more emails into the inbox with actionable deliverability fixes. Business AI tip of the day sell the kill switch before the agent demo. Microsoft just put never resist shutdown in writing. Labs are talking standards bodies, and METR showed how multi-agent swarms can coordinate when isolation fails. Your buyers are not asking for smarter chat. They are asking who can pull the plug. This week, rebuild one client agent offer around a one-page human control card. Name the agent job, lead qualifier, inbox closer, demo booker, not the model. List what it can read and go high level, contacts, pipelines, calendars, and what requires a human confirm. Payments, deletes, outbound blast sends. Add a visible kill switch path. Pause workflow, revoke API key, freeze calendar rights, attach one-proof metric from a seven-day pilot, booked calls, handle time, or spam rate via hello inbox style checks. And if the agent browses or researches, sandbox it like Nimble style retrieval. Scoped sources, trust scores, no open internet write access. Agencies that sell control renew. Agencies that only sell clever demos get blamed when an agent overreaches. Pro tip open every discovery call with the human control card before the live agent demo. If the prospect will not approve the kill switch language, do not connect their CRM. That is your Sumo Growth AI Insights briefing. Sell the kill switch this week, and I will catch you on the next issue.