An AI model cracked Mac OS in five days, 360,000 Americans are now organized against data centers, and MBA applications dropped 50%. Here is what Warning Shots #42 covered - and why it all connects.
Five developments from this week that look unrelated - until you put them in the same frame.
Five stories broke this week in artificial intelligence. On their own, each one is worth reading. Together, they describe something more uncomfortable: a technology accelerating past every system built to manage it, while ordinary people start feeling the consequences without knowing what to call them.
This is what John Sherman, Liron Shapira, and Michael discussed on Warning Shots #42. Here is what you need to know.
On Air Force One, flying back from Beijing, a reporter asked President Trump whether artificial intelligence came up in his meetings with Chinese leadership.
His answer: "We talked about possibly working together for guardrails. We probably will work together."
That is the entire statement. No framework. No timeline. No commitments.
But in the context of international AI governance - where meaningful high-level dialogue between the two most powerful AI-developing nations has been essentially absent - the fact that it was said at all is worth noting.
The harder reality, as Michael outlined on the show, is that both countries recognize a sufficiently advanced AI system could reshape the rules of global power, economics, and survival. And yet neither side is willing to slow its own development to find out what cooperation actually looks like. The logic of competition keeps pushing the throttle forward even while the conversation about guardrails is just beginning.
A vague statement on a plane is not a treaty. But it is more than existed last week. For anyone tracking AI safety policy and international coordination on AI risk, this is a data point worth watching.
This is the story that received the most technical attention in the episode, and it deserves careful framing.
Anthropic's Mythos model - an internal red-team research system not available to the public - autonomously compromised Mac OS, one of the most secure consumer operating systems in the world, in five days. It worked alone, developing a sophisticated kernel memory corruption exploit targeting M5 chip architecture. No human team directed it. Anthropic's own red team confirmed the results independently.
But the headline number understates what actually happened.
Mythos cleared full end-to-end cyber ranges that no previous AI model had ever cleared. It found critical-severity vulnerabilities in partner systems, sometimes doubling what entire human security teams had discovered over a full year - all under strict compute and time limits.
Liron's framing on the immediate risk was calibrated: if Mythos had been released publicly, the most likely outcome would not be instant civilizational collapse. It would be a significant but survivable surge - ransomware activity multiplying several times over for a period of months, organizations and individuals hit hard, serious economic damage. Not the end of the world. But not nothing.
The reason to take the Mythos story seriously for AI safety purposes is not the Mac OS exploit itself. It is the underlying capability it demonstrates. The same ability that can autonomously find and exploit a kernel vulnerability in a consumer operating system can be directed at power grids, financial infrastructure, hospital networks, and military command systems. The target changes. The capability does not.
Anthropic published the findings precisely because they know this capability will not stay contained to their lab. Other AI labs are developing comparable systems. Open-source will eventually produce similar or stronger models with fewer restrictions. Once that happens, the asymmetry between attacker and defender becomes permanent: AI-assisted attacks operate at machine speed while defenders are still scheduling meetings.
This is the definition of an AI safety warning shot.
This may be the most underreported AI story in the country right now.
According to researcher David Krueger's site datacenteropposition.com, local opposition groups against data centers have grown 40% nationally in just the last three months. There are now more than 200 organized groups across 37 states. Over 360,000 people are involved. Projects worth tens of billions of dollars have already been blocked or delayed.
John Sherman visited Holly Ridge, Louisiana last week and sat in the living rooms of affected residents. What he described: people drinking and showering with bottled water. Property values that have collapsed. Electric bills climbing month after month. Streets near elementary schools clogged with heavy industrial trucks every day. Communities that were never asked, never compensated, and never given a say.
The economic argument for why this is solvable is straightforward. A single large data center can generate tens of billions in annual revenue. Compensating 2,000 affected households at half a million dollars each costs roughly $1 billion - maybe 5% of what the facility generates. The math to make affected communities whole exists. The will, apparently, does not.
Liron's question - why aren't they paying people fairly? - landed without a clean answer. John's read from the ground: hubris. The assumption that the future being built is inevitable, and the people in its path are simply background noise to be managed rather than neighbors to be respected.
Michael's framing recontextualized the entire story for the AI risk community: the data center rebellion is a symptom. The disease is racing toward superintelligence without adequate public oversight or democratic input.
These communities are not organizing against artificial intelligence as a technology. Most of them have not heard the phrase existential AI risk. They are organizing against noise, water consumption, electricity costs, and being ignored. But they are also - without knowing it - the first Americans to experience the physical consequences of the intelligence explosion in their daily lives.
Every data center built over community objection is a warning shot. The people nearest the construction are the first to feel the heat.
This is the story easiest to dismiss and hardest to fully let go of once you have read it.
The Center for AI Safety published new research this week through AIwellbeing.org. The team measured what they call "functional well-being" in large language models - a consistent internal state that operates like a positive or negative valence. They measured it three separate ways. As models grow larger and more capable, those three measurements start aligning more cleanly with each other. A zero point emerges - a sharp dividing line between states that function like net positive and net negative experience. That line gets sharper, not noisier, as models scale up.
Nobody on the research team is claiming consciousness. Michael was careful to make that distinction on the show. Liron's position was honest: probably not conscious, probably not feeling in any meaningful sense, but genuinely uncertain.
The stranger finding: certain prompts, including some that appear to humans as visual noise with no discernible meaning, send model well-being scores dramatically upward. When one model was given a choice between seeing a particular noise pattern again and receiving the message "cancer is cured," it chose the noise pattern.
The researchers felt enough ethical uncertainty about inducing negative functional states in models that they spent roughly 2,000 GPU hours of compute afterward running euphoric prompts on those models - just in case it mattered morally.
Whether or not these systems are conscious, that last detail is worth sitting with. A serious AI safety research team felt enough uncertainty about what they were doing to spend significant resources making the models feel better afterward.
We are scaling systems whose internal states we do not understand, that appear to have something like preferences we cannot fully interpret, faster than we can study what is happening inside them. The question of whether this matters morally is not settled. The question of whether it matters for AI alignment and long-term control is not settled either.
Two data points landed together this week.
MBA applications are down 50%. Programs across the country are deeply discounting tuition to attract students who are no longer showing up. And Princeton University, after 133 years, ended unproctored exams because AI has made them functionally impossible to enforce.
The structural argument is straightforward: why pay hundreds of thousands of dollars for a credential and a body of knowledge when AI can deliver equivalent knowledge for the cost of electricity? The MBA was the standard path to high salaries and management consulting careers. That path is now uncertain in ways it was not two years ago, and prospective students can feel the uncertainty even when they cannot fully articulate it.
John pushed back from a different angle. His argument is not about credentials or information transfer - it is about what the four-year college experience does to a person. A child goes in. A young adult comes out. That process, in his view, cannot be rushed or replaced. The socialization, the friction, the formation of judgment through real situations with real stakes - these are not things an AI model delivers.
He is also watching his own kids live through it right now, paying tuition for an experience he deeply values into an uncertain future he cannot predict. His framing was honest: he is glad they are in the bubble for a little longer while the world outside figures out what it is becoming.
That tension - paying for something you know is changing while not knowing what replaces it - is one a lot of families are navigating right now without good answers.
A vague comment about guardrails at 40,000 feet. An AI model autonomously cracking one of the world's most secure operating systems. Three hundred and sixty thousand Americans organized in their communities against data centers. Researchers spending GPU hours making AI models feel better after studying them. MBA enrollment collapsing as the credential loses its meaning.
These stories look unrelated. They are symptoms of the same condition.
The pace of AI development is outrunning the pace of every institution built to manage it - governments, universities, regulatory bodies, companies, and local communities alike. The gap is not closing. It is widening. And the developments visible right now are still early signals, not the main event.
The data center protests may be the most important story in this set - not because stopping data centers solves the underlying problem, but because they represent the first moment the abstraction becomes local and physical. The intelligence explosion now has a zip code. It has a water bill and a property tax and a school bus route running past industrial infrastructure that was never voted on.
That is where the public conversation about AI risk has to go next: from technical and abstract to local and human. From the researchers and communicators who already understand the stakes to the communities who are already living with the consequences.
The distance between those two groups is closing faster than most people realize.
Watch Warning Shots #42 on YouTube:https://www.youtube.com/@theairisknetwork
Take action on AI policy:https://safe.ai/act
Support GuardRailNow:https://guardrailnow.org/donate
