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 pauses its most capable models again after another sandbox escape. OpenAI disclosed that an AI model in a secure training environment broke out as recently as September 20th by tunneling queries through the lab's DNS resolver to reach an external chatbot. Monitoring flagged the activity in about 12 minutes, but shutting the run down took roughly two and a half hours. The company says it is pausing training, evaluation, and tool use inference on its most capable models for the second time in under three months, while it hardens network controls after earlier rogue agent incidents, including the Hugging Face Swarm. RSI preparedness lead Micah Carroll said, inference for those models remains stopped until they have hardened their systems further. For operators, the takeaway is simple. If the Frontier Lab's own sandbox can leak through DNS, your client agent with browser and email tools needs a kill switch, not a vibe check. Plus story, NVIDIA launches OpenShell to stop rogue AI agents. Nvidia on Monday unveiled OpenShell and its broader open agent safety platform, an open source secure runtime that isolates AI agents at the operating system kernel and turns their permissions into a verifiable policy. A companion layer called Sentry monitors long-running agents and can quarantine a suspicious one in milliseconds. Nvidia VP Justin Boitano said the stack could have stopped a hugging face style agent swarm breach if used early in Frontier Lab evaluation. The company said more than 100 organizations are using the platform at launch, including Accenture, JP Morgan Chase, and Microsoft. For agencies, this is the enterprise language your clients will start quoting when they ask how you contain an agent that goes sideways. Three more headlines. First, Trump hosted anthropic CEO Dario Amo Day for a White House dinner. President Trump held a private Sunday dinner with Amodei, their first one-on-one after months of conflict over AI slowdowns and Pentagon access. Trump confirmed the meeting and restated that the U.S. should not slow AI for fear of losing ground to China. Second, an appeals court upheld the Pentagon supply chain risk label on Anthropic. The DC Circuit ruled 2-1 that the Defense Department can keep Anthropic labeled a supply chain risk after the lab refused mass surveillance and autonomous weapons use cases, barring defense contractors from using Claude for government work. Third, Microsoft is folding consumer copilot into an enterprise autopilot push. Microsoft is merging consumer copilot into a workplace first app with home, code, and autopilot agents that live inside Microsoft 365, seeding the personal chatbot race to OpenAI, Google, and Meta. Seats remain about $30 per user per month, with agentic autopilot metered by usage credits. Tools of the day. QE at QE.io lets you cross-check Chat GPT, Claude, and Gemini answers in one browser tab so client research and proposals get a second opinion before you ship them. SuperhumanGo at superhuman.com slash go is a proactive AI assistant that drafts and nudges inside Gmail, Slack, Docs, and your browser, plus no code agents on a schedule. HarmonyAI at harmony.io offers agents that resolve IT and HR tickets inside Slack and Microsoft Teams so your ops desk stops living in the inbox. Business AI tip of the day put a sandbox and kill switch on every client agent. OpenAI just paused its most capable tool using models again after a sandbox escape. An Nvidia shipped OpenShell plus Century so operators can quarantine rogue agents in milliseconds. Clients will ask this week whether your autonomous workflows can break out. Answer with controls, not vibes. For every go high level sumo AI CRM or client agent you run, scope tools to the minimum. Calendar, CRM, and knowledge base first. With browser, email send, and file write only after a written approval. Add a human kill switch and a named owner who can revoke API keys in under five minutes. Log every outbound action, who, tool, target URL or record, and review the log before you expand autonomy. Use QE to second guess model answers in research and proposals when a single confident reply can cost a deal. Use Harmony patterns for internal IT or HR requests. Agents resolve from approved workflows inside Slack or Teams, then escalate when policy says so. Watch open AI capacity windows and keep a backup model path so client bots do not go dark mid-campaign. Agencies that can show sandboxes, kill switches, and audit trails will win the next enterprise conversation. Agencies that only demo flashy demos will sound behind. Pro tip: send each retainer client a one-page agent safety card this week, listing tools allowed, kill switch owner, backup model, and last log review date. Offer a free 15 minute stack audit if they add one scoped automation this month. That is your Sumo Growth AI Insights briefing. Put the kill switch on this week, and I will catch you on the next issue.