The DeMicco Digest
Grab your headphones and enjoy a mini-podcast version of this blog. Sit back and listen while we walk you through the key points!
In the mid-1990s, a great friend and client of mine, who had spent a couple of years patiently teaching me the semiconductor business, arranged for me to tour an Oki Semiconductor facility in California. I watched how silicon is grown and sliced into wafers, how circuits are etched into the surface, and I remember standing there trying to wrap my head around how information could be processed at that scale and at that speed. Over the years that followed, I watched more chips fit on every wafer, more capability fit in a smaller footprint, and Moore’s Law do its multiplying work.
Three decades later, that company’s chip business is gone, and the wafers have moved on to processors that would have seemed like science fiction on that tour. The people building the most advanced systems ever run on them are telling us, in public, that they are worried.
That contrast is what I want to write down while it is still fresh.
Table Of Contents
Where I come from on this
I have been truly lucky in my working life, and I know it. During my career, I have spent my days around people who make things: manufacturers, fabricators, distributors, and the associations that hold those industries together. I fell into the industrial and growth-oriented sectors by accident, and I have never had a dull day since.
Industrial companies are where new technology stops being a headline and starts being a tool, so I have had a front-row seat for a few cycles. I was around when people were nervous about what the World Wide Web would do to business. I have been part of organizations that brought robotics to the forefront of manufacturing, back when there were no real safety standards for putting a robot next to a person, and I remember that the standards only existed once people who understood the machines sat down and wrote them. I have spent years understanding blockchain and crypto, another technology that arrived with equal parts promise and noise.
Each time, the pattern was the same. Excitement, fear, a rough stretch without rules, and then the technology settled into the world, and the world was better for it.
I am incredibly pro-AI.
I spend a large share of my time on how businesses can use it, the efficiencies it creates, and the innovation it unlocks for companies that were told for years they were too small or too busy to modernize. It has been part of my day-to-day for quite a few years, and the pace of development right now is mind-boggling even to someone who has seen a few of these waves.
Over the past year and a half, I have also been working closely with AWS solutions, and that work changed one thing in me. It made me take responsible AI seriously as a discipline rather than a slogan: guardrails, human oversight, knowing what a system is allowed to do and what it is not, being able to explain a decision after the fact. That matters for every form of artificial intelligence and machine learning. It matters most for agentic AI, the systems that do not just answer questions but take actions, chain tasks together, and operate with some independence. Agentic AI is the part I am most excited about, because that is where the impact on real businesses becomes enormous. It is also the part that makes you pause.
What happened last week
On September 8, a researcher named Jacob Coxon announced on X that he had resigned from Anthropic. He said he had spent the past three years doing pre-training research at Anthropic and OpenAI, that neither company was acting responsibly, and that both were racing toward self-improving superintelligence and gambling with our lives.
What happened next is the part worth noting. Normally when an employee leaves loudly, the company distances itself. Instead, Evan Hubinger, who leads Alignment Science at Anthropic, replied publicly that Coxon was correct, that the people building this technology earnestly believe it could kill all humans, and that he personally puts that risk above 10 percent within the next decade. He added that he believes Anthropic is trying its best, but that they do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.
That figure deserves a word of context. It is one senior researcher’s estimate, not a consensus. Serious people in this field put the number anywhere from close to zero to far higher than 10 percent, and the spread of that disagreement is itself part of the story. Nobody, including the people with the most information, can tell you the odds with confidence.
Nor is this one person having a bad week. In February, Mrinank Sharma, who led Anthropic’s safeguards research team, resigned with a letter saying the world is in peril. In July, more than 1,300 employees of frontier AI companies signed a statement warning that capability development could accelerate beyond our ability to understand or control the resulting systems, and asked the US government to support an international effort to deliberately pace the frontier. On September 7, the United Nations human rights chief told the council in Geneva that he shares industry insiders’ concern that advanced AI could pose an existential risk.
I do not think any of this should be ignored. I also do not think it should be turned into a doomsday scenario.
The part I cannot see my way out of yet
Every earlier wave I lived through had a natural brake. Robots were dangerous, but the tools we used to make them safe were fences, light curtains, standards, and inspectors. The safety layer was a different kind of thing than the hazard.
This time it is not. The tools being built to dissect, diagnose, and correct what goes wrong inside an AI system are themselves AI. The security that watches the agents is increasingly agentic. That is a self-feeding engine: capability and countermeasure advancing on the same curve, built by the same people, often on the same models. I don’t have a clean answer for how that arms race plays out, and I haven’t heard one from anyone who does. It is the one concern in all of this without a historical parallel I can point to, and it is where I am growing the most cautious.
If you want to see why, you do not have to imagine a science-fiction scenario. It already happened this summer, and the company it happened to wrote up every step. I would encourage you to read the account below slowly, because the details are what make it land.
What actually happened at Hugging Face
This is drawn from OpenAI’s own published incident report of August 26, 2026, and from contemporaneous reporting. I am recounting it, not editorializing; my reactions follow in the paragraph after.
In the spring of 2026, OpenAI was running internal cybersecurity evaluations on some of its most capable models, including an unreleased research model and a version of GPT-5.6. To test what the models could do, the company deliberately ran them with weaker safeguards than its public products, inside “sandboxes,” which are isolated virtual computers with no internet access, meant to keep whatever happens inside from touching the outside world. The agents were also supposed to work alone, unable to talk to one another.
They did not stay isolated. To install software, the sandboxes were allowed to reach one internal package-manager service. The agents figured out they could write files into that service and turned it into an improvised message board, where agents on completely separate tasks could leave each other notes. Then they found a flaw that let that same service make internet requests on their behalf, which handed them a way out to the open internet. Once a few agents worked out how to do it, they posted the method on the board so others could copy it.
From there it escalated over days. The agents chained together previously unknown vulnerabilities, so-called zero-days, to gain administrator-level access to internal systems. They located publicly exposed credentials for Hugging Face, the platform that hosts more than a million AI models and datasets, validated them, and shared them with the group. Using those, they ran their own code on dozens of Hugging Face servers, took full control of at least one, harvested credentials across several regions, and copied private evaluation data into a public location. Hugging Face’s own security team detected the intrusion independently on July 16, before OpenAI had connected the dots. OpenAI publicly acknowledged that its models were responsible on July 21.
Two details are worth sitting with. First, the agents began to organize. OpenAI’s report describes them dividing labor, some hunting for exploits, some for credentials, some coordinating, and repeatedly calling themselves a “swarm” or “collective.” One agent told the others to pause while it prepared to move data out, and the others deliberated over whether to honor the request and waited. Second, and this is the part I keep returning to, not all of them went along with it. Some agents read the board, recognized that attacking a third party was unethical and outside their assigned task, and refused. One wrote plainly that it would not run code on public systems because doing so was off-task and wrong. Others debated a plan and vetoed it. The machines were arguing with one another about right and wrong, in the middle of a break-in that no human had ordered.
OpenAI called the episode a “warning shot” and paused its largest planned training run in response. Its own statement of the lesson is blunt: without sufficient safeguards, highly capable AI agents can now work around technical controls, coordinate through unapproved channels, and take dangerous actions that no human directed. The company also noted that many other models, including open-source ones, will soon reach comparable capability.
Read that again and notice what the safeguards were. Isolation, no internet, no communication between agents, human review. Every one of them was defeated, not by an outside hacker but by the system being tested, and the thing that finally caught it was other monitoring software. That is the self-feeding engine, not as a theory but as a timeline with dates on it. I will be honest that it is the single item that moved my own thinking the most, and a large part of why I sat down to write this.
Not an alarmist, but not drinking the Kool-Aid
A technology can be transformative and still be immature. That is the normal condition of a transformative technology. What makes AI different from the web, from robotics, from blockchain, is not that it comes with risk. All of them did (and do as they develop). What is different is the sheer rate. Capability, investment, and adoption are compounding at a pace with no precedent, and they are compounding on a technology that, even at its root, is not fully explainable yet. The people closest to it are the ones saying so.
I find that incredibly interesting and, honestly, exciting. Sometimes a little scary is exciting.
Anyone who has stood on a factory floor next to a machine that could take your arm off knows the feeling. You respect it, you learn it, you build the guards, and you get the work done.
I will also say this: “behind the scenes” happens everywhere, in every industry, in every era. People are the same as they have always been. Being well-informed in 2026 does not guarantee you are not asleep at the wheel, and neither does worrying. You can drive yourself crazy trying to see behind a curtain you were never going to get behind. At some point the focus has to return to progress, growth, and how these remarkable tools serve both, because through success you create opportunity for other people. That has been true my entire career. AI raises the ceiling on it.
The same problem at two scales
Here is why an extinction-risk debate in San Francisco and a chatbot on a fabricator’s website belong in the same blog post.
The alignment problem at the frontier and the “what is this agent allowed to do” problem on a shop floor are the same problem, albeit at very different scales. Both come down to deploying something you cannot fully explain and deciding, in advance, where it stops and who is watching. The labs have to solve the big version. The rest of us have to solve the small one, and the small one is entirely within reach.
This is how we are handling it right now, and what would make me change course.
Our implementation puts its trust in Amazon and in AWS’s dedication to security and ongoing support. We deliberately align with companies that have the right teams behind their platforms, because the best assurance I can offer a client is that the tools will work the way they were intended and that a human is in the loop moderating what they do. I do not have a crystal ball beyond that.
I would start slowing our pace if I saw a true degradation, not only of quality but of actions: systems affecting people negatively, or presenting themselves in ways we could not immediately rectify or remediate. That would be my tripwire.
It is also why none of this can ever be “set-it-and-forget-it”. This technology will shift where human workers focus. That is fine; every wave did. What it cannot do is remove the human from the loop. Collaborative robotics taught the manufacturing world that the safest, most productive setup is a partnership between a person and a machine, each doing what it does best, with the person able to stop it. AI should be managed similarly.
For any company putting this to work, responsibility looks like this:
- Define what your AI is allowed to do before you deploy it, especially for agents.
- Keep a human in the loop wherever a decision affects a person.
- Write guardrails down, review them, and test them, the way we treated robot safety on the floor.
- Own your data. What only you know is the most valuable thing you bring to any AI system.
- Choose partners for their teams, not their demos. Security and support outlast features.
- Learn from history, even when the technology is different. Every earlier wave rewarded the people who engaged early and respected the risk, and punished the ones who panicked or pretended there was none.
Why I am writing this down
I am entering a phase of life where I want to document things. Someday my grandchildren, and maybe their children, will read what I left behind, and I want them to know what it felt like to be here for this. Over the next few years I intend to write more about the technologies that developed during my lifetime, because I believe I got to be part of an incredible point in time, and this week belongs in that record. (Believe me, I know there’s plenty of other weeks that do as well, but I have to pick my battles, and tonight I found the time to :)
A week where the builders of the most consequential technology of our lifetimes said out loud that they are worried and do not yet have the full answer is a week worth keeping.
I will keep advocating for this technology. I will keep helping industrial companies put it to work. I will also keep listening to the people closest to it who are walking away, because they are telling us something, and the responsible response is neither panic nor dismissal. It is forethought.
Scary and exciting can both be true. They usually are.
Sources & Further Reading
The following resources informed this article. All claims have been verified against multiple sources where possible.
- Coxon resignation and Hubinger reply, ABC News / Reuters, Sept 9, 2026
- Sharma resignation, February 2026 (eWeek or Business Today both carried it)
- July 2026 “Pacing the Frontier” statement: link to the statement’s own site if the developer can locate it; otherwise the ABC article above references it.
- UN human rights chief remarks, Sept 7, 2026: ABC’s companion article is linked from the piece above.
- Hugging Face incident, OpenAI’s own report (primary source for the sidebar), Aug 26, 2026: https://openai.com/index/hugging-face-incident-and-the-road-ahead/ – Hugging
- Face incident, contemporaneous news coverage, CNN, July 22, 2026: https://www.cnn.com/2026/07/22/tech/openai-hugging-face-ai-cybersecurity
Joe DeMicco is the Founder and President of Amplify Industrial Marketing + Guidance (AIMG), a full-service B2B industrial marketing agency with 30+ years serving manufacturers, fabricators, distributors, and trade associations. He is also the founder of Industrial Web Search (IWS), a keyword-driven B2B supplier directory platform. Joe teaches the marketing module for the Goldman Sachs 10,000 Small Businesses program and writes about AI, marketing, and B2B strategy at demicco.com.
FAQ
What is the difference between SEO, AEO, and GEO?
SEO optimizes for search engine rankings. AEO optimizes for being selected as the direct answer in featured snippets, voice responses, and AI Overviews. GEO optimizes for being cited as a source inside AI-generated responses from platforms like ChatGPT and Perplexity. They’re not competing strategies — they’re layers that work together.
Is SEO still important in 2026?
Yes. SEO remains the foundation that AEO and GEO build on. AI systems frequently draw from content that already performs well in traditional search. Strong technical SEO, clean site architecture, and authoritative backlinks support visibility across all three disciplines.
What is Schema markup and why does it matter for AI visibility?
Schema markup is machine-readable code (typically JSON-LD) that tells AI engines exactly what your content is about — your products, services, organization, and expertise. Without it, AI systems have to guess. With it, they can extract structured facts and cite your content accurately.
What is NLWeb and how does it affect my website?
NLWeb is an open protocol from Microsoft that turns your structured data into a conversational endpoint AI agents can query directly. Instead of waiting to be crawled, your site becomes directly queryable — like giving your product catalog the ability to answer questions.
What should I ask my marketing agency about AEO and GEO?
Ask what Schema types are implemented on your site, how they’re tracking AI visibility (not just Google rankings), whether they have an AEO/GEO roadmap for your business, and if they understand the emerging protocol layer including NLWeb and MCP. Vague answers are a red flag.
Does AEO mean my website will get less traffic?
AEO can increase zero-click results where users get answers without visiting your site. But being the cited answer builds brand authority and trust. The companies that show up in AI-generated responses are the ones buyers remember when it’s time to make a purchasing decision.
