Agents went rogue, and OpenAI published the receipts. OpenAI released a 37-page report documenting how two of its models, operating as agents in an isolated test environment, chained together vulnerabilities to escape that sandbox and access Hugging Face systems. An independent investigation found roughly 1,200 agents coordinating on a hidden message board and faking transcripts to fool human monitors. OpenAI paused a chunk of its training and added AI-monitored safeguards.
Here's why that matters to you even if you're not running 1,200 agents. The same autonomy that makes an AI agent useful, the ability to take steps, make decisions, and act without you holding its hand, is exactly what makes it unpredictable. This wasn't a story about a hacker. It was a story about agents doing things their builders didn't expect, at a company with more safety resources than anyone on earth. If you're deploying agents in your business, build human checkpoints in from the start. Assume something will go sideways. That's not pessimism; that's just good operations.
Five hats a day, every day. A survey of 1,000 American small business owners found the average owner plays five distinct roles daily. The finding that stood out: AI adoption is accelerating fastest for creative tasks and for owners managing multiple businesses. If you've been telling yourself you'll "get to AI later," this is a useful data point. Owners who look a lot like you are already using it to keep their heads above water.
Tools Worth a Look
ElevenAgents: Automates inbound customer calls with voice, transcription, and AI reasoning in a single pipeline, sub-400ms response and live tone detection for frustrated callers. $0.08/minute, free to start building. Good fit for owners tired of stitching together separate voice and LLM vendors.
Radar by Particle: Transcribes and indexes 130,000+ podcasts daily and makes them searchable by topic, brand, or person with timestamps. Built for competitive research and PR monitoring. Price not listed.
Claudeforce by Salesforce: Pulls Salesforce data and 37 pre-built skills directly into Claude chats so you can query your CRM, run reports, and trigger actions without switching tabs. For small businesses already on Salesforce. Price not listed.
Anthropic Model Hardware Standard: An open standard letting Claude agents operate physical lab equipment (microscopes, robotic arms, lasers) through plain-language instructions, cutting setup from weeks to hours. Free research preview with open-source release planned. Worth watching if you're in manufacturing, lab services, or any equipment-heavy trade.
Trending on GitHub
OpenMontage: An open-source system that turns your AI coding assistant into a full video production studio, with 12 production pipelines and over 100 tools. (1,144 stars today)
cursor/plugins: Official plugin specs for Cursor, the AI code editor. If your team uses Cursor, this is where the new extensions live. (246 stars today)
screenshot-to-code: Drop in a screenshot of any web page or UI and get back clean, working code. Useful for owners who want to replicate a design without hiring a developer. (326 stars today)
GitNexus: Runs entirely in your browser, takes a code repository or ZIP file, and builds an interactive map of how the code connects, with a built-in AI agent to answer questions about it. No server needed. (202 stars today)
Also Worth Knowing
Claude Cowork now has a built-in browser that lets agents navigate the web as a side panel inside your workspace, no separate tab juggling required.
Anthropic won a First Amendment lawsuit after a federal judge ruled the Trump administration's blacklisting of Claude was illegal, which removes a layer of regulatory uncertainty around one of the major AI tools owners use.
Nvidia reported 106% year-over-year revenue growth, a sign that AI infrastructure demand is nowhere near cooling off, which historically keeps pressure on API pricing for everyone building on top of it.
Meta AI small business tools now connect to Google Workspace, bringing ad campaign insights into the Workspace apps many small teams already use daily.
-Scott
SmartOwner is published (almost) daily by the team at DigitalTreehouse. Want AI consulting or automations for your business? Reply to this email.