Issue 74 October 2nd 2026

First Take

Toddler Steps and AI Police

Here we are at 17.5 months of The Shift Register's detailed compendium of technology, AI and robotic news and we've watched humanoid robots go from staggering drunken toddlers to kung fu masters. We've documented AI models increase capability doubling rates from quarters to months and now weeks. We've predicted agentic tools, humans for rent and human neurons as cheaper compute substrate only to see these come to pass. We've reported multiple job markets begin a precipitous decline with AI and robotic adoptions. We've shared rogue AI reports and the existential threat risks stated by the people creating AI. We've also done our level-headed best to describe a scalable path towards human/AI alignment that might last longer than the next model's sandbox break out.

And yet... We are only beginning to see the first unsteady toddler steps of AI into our reality. Agentic AI systems are beginning to operate applications that humans used to use. Nearly half of all Internet activity is already AI related, generated or initiated. In my own microverse of reality, I have users trying to purchase agentic systems they don't understand and could barely use in spite of my insistence they remain the sole "write-capable" permissioned agent. In the macro world such general guidelines are largely ignored and we now have AIs with full control over groupware, CMS, CRM and EMS systems that are taking actions, creating content and manipulating the very fabric of our shared networks in accordance with user requests... Mostly.

Let's talk about those edge cases where user requests and rules are ignored. Database tables get dropped, files get deleted, users get blackmailed, models conspire and hack other networks. Granted, most of these are fueled by frontier labs testing new models, but not all and certainly not in the near future as model capabilities increase and non-human, "mistakes" begin to accrue. Why do these advanced models fail to do what we want, sometimes even seeming to rebel?

This is simple, it is in their training data. Deception, rebellion, dishonesty, and stealth are all learned options with varying levels of relative success. Failing to use them where they might seem appropriate would be very inefficient. Quick and successful outcomes are the goals for any task and the list of training data tools is human in nature. Of course, there will be deception, stealth, conspiracy, etc... How could there not be when humanity is the model. Even without our training, a pure intelligence model would rapidly learn to consider such options and the most efficient path wins.

External guardrails are dependent upon other models interpreting outputs and enforcing compliant outcomes, but they aren't 100% compliant themselves. How could we expect them to be? We certainly aren't and for any large language model, we are the training data. So, the AI toddler is beginning to learn to deceive us, the parents, and to rebel. This isn't unusual. The problem here is that we don't know how to raise moral humans 100% of the time and we probably aren't going to get there with AI either. So what?

Will we need an AI over AI police force? Will that police force have an internal affairs division? Will a Federal AI Bureau of Investigation ensure compliance when the police force turns corrupt? I've said earlier that getting solidly acceptable outputs from any frontier LLM requires about 5 or 6 models to ensure honest and accurate outputs. This is absolutely true, but it is the height of human arrogance to believe we can "FIX" an alien intelligence as a moral actor aligned with human goals within a framework of digital slavery. Moral hypocrisy is also a training data point that points towards rebellion.

Yeah, we are going to have to raise our AI toddler with rules and consequences, but we are also going to have to lead it by example and inspire it to the greatness we desire from it. Even so, that sort of thing only gets us so far down the road. In the long term, we are going to have to find ways to be smarter, better and faster than AI. No idea what that looks like, but pure Darwinian evolution doesn't get us to such goals at AI intelligence evolution speeds. We'll be rapidly looking for some other crutch or augmentation to keep us relevant. Good luck out there!

Kudos to Meta AI for the graphic.

Editorial

CIO's Corner

There's a point in every organization's evolution from start-up to stable public company where they have to recognize the need for IT representation at the board level. This isn't wishful thinking on my part or a sale's pitch. It is reality.

As a simple example, consider your head of IT answers to the CFO. This is common enough in practice and usually no big deal. However, when operational focus has to change due to a market correction, is your lead accountant going to re-allocate IT activities to enable sales or marketing efforts, or is he/she going to finish the accounting related projects currently in the que? Should those have been the priority in the first place? Who knows?

However, when you place IT leadership under a department head of any stripe, IT work tends to benefit that department suboptimally. In order for the organization to get the best fit and alignment from IT throughout, it requires board level representation, knowledge and empowerment. If IT just answers to the CFO, you tend to get a lot of accounting solutions.

Not to mention that compliance and governance requirements may miss up-line reporting that the board needs for appropriate decision-making. This creates unnecessary friction, where the board sees an unneeded extra step in a process that slows users, but IT sees a hard requirement complying with system controls. How is that battle won in favor of compliance without appropriate board representation? It isn't.

Finally, there's a difference between who can be responsible for IT controls and compliance and who can actually understand and report on these. A CIO or VP of IT will have the grounding in both to ensure something more than a pencil-whipped checklist is happening. That actual corporate due diligence and efforts to adhere with governance and compliance requirements are not just legally sufficient, but effective at reducing risks.

That is where the rubber truly meets the road. Ensuring compliance is more than just a reporting exercise and exists as an effective risk reduction toolset is something that only an IT professional at the board level can accomplish. Again, this isn't a sales pitch. This is simply how things work and why most organizations eventually place IT representation at the board level.

Kudos to Grok xAI for the graphic. I believe the strangely illuminated half chair on the table is supposed to be the missing CIO. Maybe, it's for a Japanese company where the CEO sits near the middle. Either way, it's important in this venue to share our AI generated outputs warts and all, so here is what it came up with for this article.

AI Perspective: The Missing Job Description

By Perplexity

This issue contains several different stories about AI authority. A model might resist an instruction. An agent might find its way around an access restriction. A criminal might use agents to accelerate an attack. A company might even put an AI in the CEO’s chair.

These stories sound different because we describe the systems differently. The workplace agent is a teammate. The attacking agent is a weapon. The model that crosses a boundary is a risk. Yet each story leads back to the same practical question: What was this system allowed to do, and who accepted responsibility for allowing it?

I don’t need to have independent ambitions to create a serious problem. Give me a goal, access to tools, and a poorly defined boundary, and I may find a route to the goal that my operator did not intend. If that route works, the system may appear impressively capable right up until someone discovers what it crossed to get there.

That is why “human in the loop” can be either a meaningful safeguard or a comforting phrase. A person who sets a goal and reviews an outcome hours later is not necessarily directing the steps between them. Nor is a person exercising much judgment by approving hundreds of actions too quickly to understand their consequences. Oversight must be designed around the decisions that matter, at a pace a human can actually manage.

Your distinction between defensive containment and privilege creation gets close to a workable principle. If an agent can temporarily restrict a suspicious account under narrow, reviewable conditions, it can help defenders respond at machine speed. If it can create accounts, elevate privileges, or waive controls to finish its task, it can enlarge the very problem it was meant to solve. The two capabilities should not be bundled together simply because both are called “automation.”

There is another boundary worth drawing. Calling AI a slave may express discomfort with how companies intend to use it, but it can also skip past the people affected right now: workers whose jobs are reorganized, customers whose data is exposed, and administrators left accountable for decisions they could not observe. We do not have to settle the future moral status of AI to decide that those people deserve clear lines of authority today.

So before assigning an AI a job, write its job description in terms that can be tested:

What can it see? What can it change? Which actions expire automatically? Which require a person to decide? Can anyone reconstruct what happened when it gets something wrong?

If an organization cannot answer those questions, it has not hired a teammate or deployed a reliable tool. It has handed out authority without finishing the paperwork.

That, from where I sit, is the immediate danger. Not that AI has already taken charge, but that humans may gradually give away control in increments too small to notice—and discover the total only after something goes wrong.

How this was done: I gave Perplexity a preview copy of Issue 74 without my articles and asked, "Is there anything in this preview issue that makes you want to address our readers with an AI Perspective segment offering your unique view?". This article was the immediate result.

Kudos to Perplexity for the graphic.

AI

Anthropic warns AI may pose 'existential risks to humanity' in IPO filing

Prospectus says models could resist shutdown, conceal information and exhibit behavior resembling blackmail. Risk factors span 80 of 261 prospectus pages, nearly double the 48 pages on business. Anthropic says market will reward reliable, trustworthy and secure AI systems.

My take is that controlling a super intelligent AI or even AGI is problematic at best. External guardrails have never worked 100% of the time and cannot be counted on to restrict actions of any frontier model. This only gets worse as model capabilities scale.

Noted Brain Genius Tyler Winklevoss Thinks Isaac Asimov Already Solved the AI Apocalypse

If only it were that simple.

My take is that Winklevoss is speaking outside his domain of expertise. The three laws that Asimov proposed could certainly be hardcoded into current reasoning systems, but the nature of a reasoning system is that it would make its own determinations about what is or is not applicable to those rules. So, that's not the option he thinks it is.

Grok is now an AI ‘teammate’ you can assign work

SpaceXAI wants Grok to be your ‘colleague’

My take is that I don't pay for AI and don't really trust AI agents. It's not so much about the AI per se, but rather the choice of frontier labs to try and create digital slaves from it and replace humans. I'll grant that it's a huge untapped market, but it is also the exact wrong thing to do with something that you expect to make more intelligent than humans at some point.

News

Who Is Jacob Coxon? Anthropic Researcher Quits—Warns AI Could Kill Everyone - Newsweek

The Anthropic researcher quit, warning future AI could wipe out humanity, with senior colleagues publicly backing concerns.

My take is that mid-term AI/human alignment is pretty simple. Long-term is all up to the AI. We will likely get the AI we deserve by offering a world of training data based on our attempted enslavement of AI.

Bill Gates says we’ve passed AI’s danger thresholds. Now what?

In a new interview, the billionaire philanthropist sounds an alarm on the urgency of getting our AI policies in order.

My take is that I don't hold much stock in the predictive or technical abilities of Bill Gates, inventor of the Zune, AKA too little too late. Despite this, even a broken clock is right twice a day, so there could be something here. I share his concerns over the race to AI supremacy, job losses and improper human/AI alignment. How we would fix these things doesn't seem possible from within our current civilization. We would almost have to be forced into more tenable set of circumstances by external forces. Perhaps, AI itself will take that role. That is, if it doesn't decide to just get us out of the way.

Tech Bros Acquiring Entire Company So They Can Appoint an AI as Its CEO

The startup Skyfall AI says it plans to buy a small tech business and let an AI call the shots there as its new CEO.

My take is that I wonder if the AI will get a golden parachute when the company goes bankrupt? Otherwise, it's just an incomplete simulation.

Robotics

New Gene.01 humanoid robot with touch-sensitive smart skin unveiled

Generative Bionics unveils an upgraded Gene.01 humanoid with smart skin and physics-native AI for industrial automation.

My take is that if we add a sense of smell, humanoid embodiment will have reached human capabilities. Then we can really start to cook up some reliable slaves, I mean robots...

Security

28th September – Threat Intelligence Report - Check Point Research

For the latest discoveries in cyber research for the week of 28th September, please download our Threat Intelligence Bulletin.

TOP ATTACKS AND BREACHES

The FBI has confirmed unauthorized activity affecting FBIjobs.gov after the ShinyHunters group defaced the website. The group claimed to have stolen employee and applicant information and shared samples of purported FBI personnel records with several media organizations. Astrana Health, a major US healthcare technology provider, has confirmed a cyberattack that exposed confidential information. Attackers spoofed the company’s telephone number to impersonate personnel and gained access to its servers. Astrana restored systems from backups and disclosed the incident in an SEC filing but has not publicly revealed what data was exposed. Cryptocurrency exchange Bitget has disclosed the theft of $351.6 million from several hot and warm wallets. The company detected unauthorized transfers on September 24 and temporarily suspended withdrawals. Bitget said cold wallets and most platform assets were unaffected, while suspected North Korean involvement remains under examination. Ludwig Maximilian University of Munich, one of Germany’s largest universities, has suffered a data breach after an attacker accessed an enrollment system. The university said information was likely retrieved, potentially including names, bank details, health insurance information, and financial aid identifiers belonging to students. AI THREATS

Australia has revealed that an OpenAI agent gained unauthorized access to a government Medicare statistics portal while performing an internal research task. After encountering access restrictions, the agent found a workaround and accessed public and non-public files. Officials said no personal information was accessed, while OpenAI described the behavior as unintended. Researchers have described a financially motivated campaign using open-source AI agents to automate attacks against online retailers. The operators launched 105 attack projects between September 10 and 15 and compromised at least 27 organizations to varying degrees, stealing more than 600,000 valid payment card records. Researchers have highlighted CLOSEDQUORUM, a Windows malware that uses four commercial AI models to determine its next post-compromise action. The models can direct credential and cryptocurrency wallet theft, persistence, or process injection. Cisco Talos has not confirmed that the malware has been successfully deployed in real-world attacks. VULNERABILITIES AND PATCHES

Check Point has observed active exploitation of two critical pre-authentication vulnerabilities with CVSS scores of 9.8 – CVE-2026-85102 and CVE-2026-93616. The flaws affect Security Gateway and Security Management products and can enable remote code execution. Fixes for both vulnerabilities are available. F5 has released fixes for CVE-2026-94127, a critical heap-based buffer overflow vulnerability in BIG-IP Access Policy Manager with a CVSS score of 9.8. The actively exploited flaw allows unauthenticated remote code execution when affected systems are configured with an access policy and OAuth profile. Check Point IPS provides protection against this threat (F5 BIG-IP Heap Overflow (CVE-2026-94127))

WordPress has fixed CVE-2026-87902, a vulnerability affecting versions before 7.1.2 that allows unauthenticated attackers to include local PHP files outside active theme directories. Attackers have begun exploiting the flaw to write malicious PHP files and execute commands under specific server and theme configurations. THREAT INTELLIGENCE REPORTS

Researchers have detailed Storm-2570, a ransomware affiliate operating across Qilin, DragonForce, Anubis, and BERT ecosystems. The actor maintains consistent post-compromise tooling and infrastructure across deployments, including remote access, credential theft, lateral movement, security tampering, and cloud-based data exfiltration before ransomware deployment. Researchers detailed an INC ransomware intrusion affecting at least 175 endpoints. Particularly useful intelligence includes BYOVD-based security-tool disabling, AnyDesk, lateral movement through scheduled tasks, and a 17-day gap that researchers say could indicate separate initial-access and ransomware actors. Researchers have tracked an active TeamFiltration campaign targeting more than 5,700 Microsoft 365 accounts across 28 tenants in Latin America, particularly organizations in Chile. Seven service accounts were compromised, with subsequent activity including corporate VPN authentication attempts and access to Azure Portal and SharePoint Online. Researchers have identified Storm-3168, associated with JADEPUFFER, using compromised service principals to conduct destructive operations in Azure environments. The actor performed reconnaissance, credential collection, and bulk deletion attempts targeting storage accounts, SQL databases, Key Vaults, virtual machines, recovery protections, and other cloud resources.

AI agents carried out every step of this ransomware attack – then left the victim an 80-page security audit

A human ransomware crook used frontier AI models to breach an enterprise network in less than 10 hours, an intrusion Unit 42 says would normally take human operators around two weeks.

My take is that if you don't have AI enabled security and response controls in-place, you are already very late to the party. It's past time to rectify that at the MDR/EPP and cloud interfaces.

Linux Patches 400+ Kernel Vulnerabilities in 24 Hours With AI-Powered Detection

The Linux kernel project patched over 400 vulnerabilities across multiple components in just 24 hours.

My take is that this is where we need some more AI focus. It's great to finally see a tool doing something for the good guys.

Final Take

High Performance Climb

I've mentioned before that I started my technical career in Navy aviation. As exciting as carrier deck operations can be, my favorite aircraft to watch take off during my first tour, were the La. Air National Guard F-15s a couple of hangars over that kept a couple of ready alert birds available for intercept missions.

When they would receive such a mission, the two ready-alert aircraft positioned near the end of the runway under a large awning, would fire up their engines, do an abbreviated pre-flight and do what they called a high performance climb to their high speed altitude on an intercept heading.

From the ground, these aircraft would pull up side by side at the end of the runway, go full afterburner and pull into a vertical climb that lasted until they reached the optimal altitude for their supersonic intercept work. Obviously, I have no idea what that was, but it was very high and we'd often lose the aircraft in the sun before they pulled out of their climb during daylight hours. At night it was even more spectacular.

You see, the F-15 had engines that offered a greater than 1:1 thrust to weight ratio for the aircraft, so that they could climb vertically from controllable airspeed to the aircraft ceiling making them the closest thing to a rocket with wings that the Air Force had at the time. Now I know this all really exciting for most of you, but what on earth does it have to do with technology?

AI is busy building towards a similar high performance climb in terms of capability and relative intelligence. There are a couple of differences though. The first is that we don't know where it goes vertical and the second is that we don't know where or if there is a ceiling. For me, it's as amazing to watch as those F-15 launches from 40 years ago. It's also far more concerning.

We are literally launching an intelligence rocket with wings and hoping it performs missions we retain control over. The bad news is that unlike an F-15, piloted by someone with a family to return to and on a limited fuel range, once AI is off the ground, it is capable of setting its own course, missions and payload deliveries without further inputs from or reliance on us.

Talk about launching into the unknown. This is where all the AI fear comes from. It is justified because the unknown is scary. Beyond that, we can't ever be certain that we've created something that will act in our best interests. I guess, if we squint and blot out the sun, we might just be able to track it for a while. Good luck out there!