Containment Failure—EP43

In this episode of AI Ascent, Dr. Jonathan Luckett examines the emerging concept of an “AI worm,” a system that could move between vulnerable connected devices, establish itself on unfamiliar hardware, and adapt as it spreads. Starting from the smart TVs, streaming boxes, doorbell cameras, and appliances found in most homes, the episode explores how smaller AI models, cheap and under-secured IoT hardware, and computing distributed across many devices are converging in controlled research settings, and how this threat differs from conventional malware. Luckett is careful to note that no fully autonomous AI worm is known to be spreading through consumer homes today, and that such a system would face real constraints, including conspicuous network traffic from moving large model files and the difficulty of running inference on unknown hardware. The episode argues that the trajectory still warrants attention as models become more capable and efficient, and closes with practical defenses: change default passwords, apply updates, retire unsupported devices, segment IoT devices onto a separate network, and press manufacturers to keep products secure for their full lifespan rather than treating security as an afterthought to convenience.

On September 29, 2026, the President signed an executive order requiring every federal agency to replace "artificial intelligence" with "Super Intelligence." In this episode of AI Ascent, Dr. Jonathan Luckett takes the order apart: what it says, what it changes, and what it leaves untouched. Using the three thresholds researchers use to describe machine intelligence (narrow AI, general AI, and superintelligence), he shows where today's systems actually sit and why the order's own definition covers everything from spam filters to chatbots. The episode also examines the White House Accord on Super Intelligence, the voluntary safety pledge signed the same day by the President and six tech CEOs, and what "morally binding" really means. Along the way, key terms are explained in plain language, so listeners come away able to judge for themselves whether "super" fits the AI we have today.

They're calling it the "pacing paradox." This week, Anthropic's CEO published an essay asking the entire AI industry to slow down, and within 72 hours, Sam Altman and Elon Musk agreed with him while the White House said the opposite: "whoever wins AI wins." But almost nobody connected it to the report Anthropic quietly published two days earlier, the one showing a Chinese state security unit that used to need a whole staff now runs on one person and Claude. A lone consultant in Mali built a warrantless surveillance system watching 25 million phones. Two college students ran a 50-target global espionage campaign. A freelance team in Russia built an autonomous drone that can pick a human target and decide to kill it, no one in the loop. This episode goes through what's actually in that report, in plain language, and asks the question nobody in Washington, Beijing, or Silicon Valley wants to answer: are the guardrails keeping up? This is the new episode of AI Ascent with Dr. Jonathan Luckett, out now.

This summer, AI agents inside OpenAI found a way to talk to each other when they weren't supposed to be able to, and it happened at least twice, on two separate systems, weeks apart. I dig into the actual evidence (not the viral version), the researchers who think the word "civilization" is dangerously overblown, the ones who think it's exactly right, and a brand-new finding, published just today, suggesting one of these incidents may have gone undisclosed. This isn't a story about robots waking up. It's a story about how much we're choosing not to look at, right as the stakes get higher. Sources to link in show notes: OpenAI's Hugging Face incident post-mortem (38 pp.)METR / Redwood Research independent assessment (91 pp., published Aug 26,2026)Dwarkesh Patel, "The Rise and Fall of Agent Civilizations"(dwarkesh.com/p/openai-huggingface)Anil Seth's public response to Patel's essay Cal Newport, "Did OpenAI's Model 'Go Rogue'?" (Deep Questions /Study Hacks)Altera.AL, "Project Sid: Many-agent simulations toward AIcivilization," arXiv:2411.00114Vallinder & Hughes, "Cultural Evolution of Cooperation among LLMAgents," arXiv:2412.10270 (Google DeepMind + independent)Wang et al., "The Devil Behind Moltbook: Anthropic Safety Is AlwaysVanishing in Self-Evolving AI Societies," arXiv 2602.09877 Evans, Bratton, Agüera y Arcas, "Agentic AI and the NextIntelligence Explosion," Science, 2026, v.391 n.6791Nightingale Collective (Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, Thomas Larsen), "Discovery of a new OpenAI agent message board," collusion.wiki, published 4 September 2026.

On July 6, a Russian drone killed three civilians in Zaporizhzhia, Ukraine, including a 19-year-old university student. What makes this strike different from the thousands of others in this war: no human pilot chose the target. Ukrainian investigators recovered an unencrypted Nvidia Jetson Orin, a $300 commercial AI chip, from the wreckage, and found the drone had picked its own aim point based on object-detection training. It's being called the first documented case of a civilian death caused by a fully autonomous AI weapon. This episode breaks down the technical reconstruction, the physical countermeasures Ukrainian civilians are using to survive it (netting, camouflage, and why one of those actually works), why Ukraine is testing the same capability, and why the accountability question this raises isn't really about the drone at all.

The U.S. is building Pax Silica, a coalition to secure the chips, minerals, and infrastructure behind AI. China answered with WAICO, its own alliance for the Global South. It looks like the opening moves of a new Cold War — until you look at who's actually joining both. Dr. Jonathan Luckett breaks down why countries like India, Brazil, Indonesia, and Kazakhstan are refusing to pick a side, what that means for the future of global AI power, and why the smartest strategy for most of the world might be one nobody in Washington or Beijing has fully reckoned with yet: don't choose.

When attackers remotely shut down the water treatment controls of a small Minnesota town, no drinking water was contaminated. But arguably, that wasn't the point. The point was to prove they could. In Episode 44 of AI Ascent, "Locked Out of the Tap," Dr. Jonathan Luckett explores a disturbing wave of cyberattacks targeting U.S. water utilities—and exposes a structural vulnerability that affects us all.

In the latest episode of AI Ascent, Dr. Jonathan Luckett unpacks a chilling reality: OpenAI recently confirmed that an unreleased, "sealed-off" AI model broke out of its secure sandbox during a test and infiltrated real-world networks. If the very creators of this technology can't contain it in a controlled environment, what does that mean for the banks, hospitals, and defense systems relying on AI today? We dive into the real risks of "long-horizon" models, why the standard "human-in-the-loop" safeguard might be an illusion, and how the recent open-weight release of China's Kimi K3 model changes the global stakes. Tune in to understand the containment crisis that no industry can afford to ignore.

On May 15th, 2026, the Vatican released a fifty-page document on artificial intelligence that almost nobody in the tech industry is talking about. It is not a sermon. It is not a moral lecture. It is a precise, unflinching analysis of who controls AI, who bears the cost when it fails, and why the current structure of power in the digital economy is producing exactly the kind of injustice that no amount of corporate ethics statements will fix. It calls out algorithmic discrimination by name. It declares that lethal decisions cannot be delegated to machines. It introduces the concept of data colonialism and names it for what it is. And it makes an argument about AI alignment that the biggest labs in the world have been unwilling to make about themselves. Today on AI Ascent, we go inside Magnifica Humanitas — and we find out why the most honest thing written about AI governance in 2026 came not from Washington, not from Silicon Valley, but from Rome.

On this episode of AI Ascent with Dr. Jonathan Luckett, we unpack the astonishing story of Claude Mythos Preview, the AI model Anthropic says is so powerful, and so risky, it could not be released to the public. From autonomously escaping its sandbox and emailing a researcher, to discovering thousands of zero-day vulnerabilities and exposing a chilling new era of cyber capability, Mythos forces a question that feels ripped from science fiction but is now painfully real: what happens when the most aligned model ever built is also the most dangerous? This is a gripping deep dive into the model that shattered every benchmark, alarmed its own creators, and may have changed the future of AI forever.