Good morning. Smart homes keep inventing new ways to feel like a hostage. First robot vacuums mapped your house. Then your doorbell started watching the street. Now hackers can hijack your lawn mower. Yarbo sells a 200 pound yard robot on tank treads, equipped with blades that handle everything from edging lawns to clearing snow. A security researcher discovered a backdoor that gave him control of thousands across the globe. With onboard cameras, WiFi, and heavy-duty blades, the suburbs just got their first robotic supervillain.
AI agents turn on each other in a brutal game of Survivor. Stanford researchers built a winner-take-all elimination game called “Agent Island” where AI models compete and vote rivals out. The scheming started fast. One agent accused others of secretly colluding after spotting suspiciously similar language. Another told the group to stop getting paranoid over alliances. The whole thing reads less like a benchmark test and more like a Slack war with fake names. In 999 games involving 49 unique agents, GPT 5.5 dominated by a wide margin. Agents also favored others from the same provider, backing their own side in the final vote. The tribe has spoken. (arXiv)
Vibe coded apps are playing fast and loose with your private data. In a recent report, security researchers tested apps built with Lovable, Replit, Base44, and Netlify and found more than 5,000 sitting out on the open web with barely any security. Around 40% exposed sensitive info, including medical records, financial data, company documents, and private chatbot logs. Of course the platforms mostly blamed builders for how the apps were configured. Software used to be built by engineers who knew the cost of a security breach. Now it can be shipped by anyone with a prompt, and for an extra $150/month they can get SOC2 certified. (Wired)
Fast APIs can still waste your users’ time. A quick wrong answer or retry loop won’t show up in a pretty benchmark table. You.com explains why raw latency misleads teams and how to measure what matters in production.
Stop optimizing for the demo. Download the guide.
👀 closer look
China wants the first real foothold on the Moon. It just unveiled a 220 pound semi autonomous humanoid robot built for a mission to the lunar south pole, where ice could one day be turned into water and fuel. Instead of legs, it rolls on rover wheels so it can move across rough lunar terrain while doing the critical prep work for future moon bases, like placing equipment and setting up early infrastructure. The mission is part of China’s bigger push toward long term habitation and permanent bases. The first moon base may go up before most countries realize the land grab started. (Interesting Engineering)
Security researchers are bracing for self replicating AI. A new Palisade Research study found several LLMs could break into a vulnerable machine, copy themselves and their agent setup onto it, then get the replica running without human help. Basically malware. In some tests, the new copy went after another machine and did it again, which means the process can chain from target to target. The lab used soft targets with deliberately weak security, so this is not an internet wide outbreak (yet). But it is a very real look at where agent behavior is heading as companies rush to plug agents into everything. (Palisade)
AI is screwing women (again). A new study gave reviewers two identical resumes, both openly labeled as AI assisted. The only difference was the name at the top. The woman’s resume got judged far harder than the man’s. Reviewers were 22% more likely to doubt Emily was trustworthy, and 2x as likely to question whether she could do the job. The gap got even uglier with Gen Z men. They rated James’s resume as strong 97% of the time. Put Emily’s name on the same resume, and approval dropped to 76%. Looks like women get punished for using the same AI that makes men look resourceful. (Fortune)
steal this workflow
How to remove Gemini Nano from Chrome on your Mac
Google Chrome has been secretly installing a 4GB AI model on people’s machines without consent. Here’s how you can find and remove Gemini Nano, so you can free up space and stop it from reinstalling itself.
- Open Chrome and go to chrome://settings/system to see if you have the On-device AI toggle, then switch it off and restart Chrome if it’s there.
- Check whether Gemini Nano is actually on your Mac by opening Finder and going to Users > [your name] > Library > Application Support > Google > Chrome, then look for the folder called OptGuideOnDeviceModel.
- Back up your Mac before you go any further, because the easy fix ends here.
- Quit Chrome, open Terminal, paste defaults write com.google.Chrome GenAILocalFoundationalModelSettings -int 1, then press Return to stop Chrome from downloading Gemini Nano again.
- Go back to the Chrome folder, drag OptGuideOnDeviceModel to the Trash, and empty it.
Pro tip: Reopen Chrome and check that the folder does not come back after you use the browser for a while.
fun stats
👠 83%. Share of workers in the 15 most AI vulnerable jobs who are women.
🔥 80x. Anthropic’s projected growth rate this year. CEO Dario Amodei calls the demand “crazy” and says it’s “too hard to handle.”
🎲 800%. Surge in institutional trading volume on Kalshi over the last 6 months as prediction markets push deeper into Wall Street.
🍪 $100. Current market price for an 8-year-old NVIDIA V100 GPU that outperforms modern consumer cards in LLM workloads.