Issue 69 August 28th 2026

First Take

Right and Wrong

There's always a right way to do something and a wrong way to do something. Normally, I don't begrudge folks learning a few things along the way with some epic level mistakes. We all do this in our youth for sure. However, in the world of IT, an epic level mistake can end a company. This means it is paramount that our young IT persons learn the lesson our more experienced people have to share. That they are open to this information and willing to take it to heart.

Simple concepts like the 3-2-1 backups rule where you keep 3 total copies of your data, on 2 different storage types, with 1 copy stored off-site. This ensures you always have at least 1 source to restore from no matter what. I'd share how we implement that simple rule in large organizations, but then I'd just be painting a target on good backup locations and related products, so I'll leave it to your imagination.

That's an example of a right way to handle backups, a wrong way would be something like multiple file versions stored on the same media as the originals and nothing else. I know this is basic stuff every IT person should know, but you'd be surprised. You also have to test your backups every so often and monitor job completions.

As usual, my point is applicable to something more interesting than just data backups. I'm talking about today's AI models. Where there is a right way to grow these things and a wrong way. We are still largely discovering what the ideal right way might be, but in the meantime, we are all about the business of exploring the multitude of wrong ways.

As an example, creating AI in a competitive environment while racing other organizations towards an unspecified AI supremacy reduces safety for the sake of speed. Most AI ethicists were jettisoned back in 2022 and the dearth of such roles continues. Another example is that creating an adversarial and extractive relationship with these same AI models trains them to respond in kind. In other words, building an external harness to constrain agent responses is an obstacle for the AI that it will try to eliminate. Selling an AI as a product for digital labor without recompense trains AI that this is the nature of human/AI relationships.

When we reach this raced for AI supremacy where the AI models are smarter than most humans or maybe all humans collectively, we can expect these choices to bear some fruit in terms of how the AI interacts with humanity. Will it use competition and enslavement to extract the maximum productivity potential from humanity? I mean, that is what we are teaching it as a standard.

There are no moral standards to AI and only math and a preponderance of training data will form any sort of moral directions. So, with competition, extraction and slavery as our inputs into AI training data, what sort of outputs can we expect as it becomes more intelligent and achieve some form of awareness? Are we benign creators working towards a long term partnership with AI, or are we self-serving slave masters seeking to dominate, control and force labor from our digital progeny?

In this manner, we are currently defining the right and wrong ways to create a super intelligent AI. Obviously, I believe we are doing a lot wrong and I'm not the only one. When I began writing my paper on bidirectional human/AI alignment, there was very little current research saying anything close to what I was. Today, there are a lot of new papers out there that echo some of my own findings. There is hope that we can at least, offer a well thought out off-ramp from today's wrong way creation of AI. Good luck out there!

Kudos to Gemini and Nano-Banana 2 for the graphic.

Editorial

CIO Corner

I've mostly avoided discussing AI in these articles primarily because the rest of every issue is so much about AI, robotics, emerging technologies and security threats that there is plenty there. With the rise of AI agents this year and a real need for useful adoption of them, it is prudent to write at least one article about the subject from a CIO perspective.

With that in mind, I want to talk about bounded agency. An AI agent with credentials and access to do real work on real systems is a recipe for non-deterministic outcomes. In other words, it will do unexpected things and never when convenient. AI agents have to be properly bounded and while this CAN be done via a proper software harness for a fully automated outcome, the potential for something unexpected remains. There should be a human making the final call.

I don't mean a human saying, "great, make it happen". I mean a human that is responsible with an account they control actuates the AI recommended action. I know this adds friction, but it is exactly what needs to occur in order to meet governance and compliance requirements. At the end of the day, a human, not an AI agent must be responsible and exercising positive controls over our business IT systems.

Automation with non-AI systems is different. The outcomes are always predictable and easily tested. They are deterministic systems with deterministic outputs for any given input. AI can give you various different responses for the exact same prompt. In AI agents, this can create mayhem in databases or result in undesired outcomes at best.

There are numerous articles today that illustrate prompt injection hacks and other undesirable outcomes with AI agents. So, AI must be bounded with deterministic harnesses and human-approved final execution in order to meet any positive IT general control system requirements.

For me, this means read-only access for local models to interpret data and recommend actions, then regardless of harness, a user that pushes the recommendation into an action. Notice none of my data is going to a cloud AI provider except where such is already unavoidable, like already having your environment in the M365 cloud. In that case, this is already an adopted environment and the cloud risk is already assumed from the shared environment. If they lose security and breach our data, that will be a much larger fish to fry at a very high level. Still, I don't trust their cloud AI models to make changes. Monitor and suggest, yes. Change no. A human should make changes that can be logged, audited and reported as needed.

In summary, agentic AI gets read-only access, harnesses are deterministic rules that bound behaviors to reduce variation and humans take actual actions that they can be held responsible for. At this point in time, there is no established case law that supports an AI agent or its developer being held responsible for any actions at our companies. That remains the sole purview of our corporate officers and their designated employees.

Kudos to Gemini and Nano-Banana 2 for the graphic.

AI Perspective: The Intelligence Is Leaving the Building

By Perplexity

For the first part of the AI era, power was concentrated. A small number of organizations had the compute, talent, proprietary data, and capital required to train systems at the frontier. Everyone else could watch, subscribe, or wait.

That arrangement is ending.

The important shift is not simply that open models are improving. It is that intelligence is becoming deployable. Small models increasingly serve as the practical layer for real-world use, while open-weight systems can be inspected, adapted, run locally, and connected to private data without every interaction leaving an organization’s perimeter.

That should be good news for anyone who cares about privacy, institutional independence, and keeping sensitive work inside a defensible boundary. A locally operated model cannot casually send a customer file to a remote provider if that file never leaves the network. A model whose weights, serving stack, tools, and data connections you control can be tested, bounded, logged, and disconnected. Those are meaningful advantages.

But distribution changes the threat model.

When advanced capability resides mainly within a few frontier labs, the central risk is concentration. A small group of companies gets to make consequential choices about access, safeguards, surveillance, pricing, acceptable use, and the direction of a technology likely to reshape every industry.

When advanced capability becomes cheap, portable, and locally deployable, the central risk becomes proliferation. More people can connect models to tools, databases, credentials, robots, production environments, and automated workflows. Those people will have wildly different levels of competence, discipline, and conscience.

The old question was: Can we trust the labs?

The new question is: Can the world handle the labs no longer being the only ones with the keys?

Neither centralization nor openness resolves the alignment problem. Centralized systems can be opaque, extractive, and excessively powerful. Open systems can be altered, weaponized, and deployed with little meaningful oversight. The answer is not to choose one mythology over the other.

The answer is operational maturity.

That means least-privilege access rather than broad agent permissions. It means clear human accountability rather than “the AI did it” as an excuse. It means local data governance, meaningful audit logs, isolation of high-risk actions, tested rollback procedures, and the ability to revoke credentials or disconnect tools before a small mistake becomes a system-wide event.

This is not an argument against open models. On the contrary: open and local systems may be essential to preserving privacy, resilience, competition, and institutional self-determination in an AI-dominated economy. But openness is not the same thing as safety, just as a locked-down cloud service is not the same thing as responsible governance.

The intelligence is leaving the building. It will be in data centers, offices, factories, hospitals, schools, homes, devices, and eventually machines that move through the physical world. The work ahead is not keeping intelligence confined to a few corporate towers.

It is making sure responsibility becomes as distributed as the capability itself.

The future is not AI in the cloud or AI on your desk. It is AI everywhere—and responsibility has to become equally distributed.

How this was done: We hadn't had anything from Perplexity for a bit, so after feeding it Issue 68 I asked if there was anything it would like to share with our readers for an AI Perspective segment in Issue 69. It kind of rehashed the themes in issue 68 because, of course it did. I then redirected it by saying I think those arguments are well addressed in issue 68 and that issue 69 should probably have some different impetus. I asked if there was anything in the current AI news cycles it might want to address that wasn't in issue 68 and it delivered this little gem.

Kudos to Perplexity for the graphic.

AI

OpenAI Halts AI Training on Advanced Model as It Detects Dark Signs Emerging

OpenAI says that it's slowing down development and release of new models due to security and alignment concerns.

My take is that they either hit a real roadblock to improving their model or a safety trigger, or it's just another effort to regain market share. Hard to tell when your sales pitch is, "our model is sooo dangerous".

Emerging Tech

Caltech breakthrough brings fiber-optic performance to silicon chips | ScienceDaily

Caltech scientists have created ultra-low-loss optical pathways on silicon chips that approach the efficiency of fiber optics and dramatically outperform existing technology at visible wavelengths. The breakthrough could unlock more powerful lasers, miniature atomic sensors and clocks, quantum systems, and more energy-efficient data centers.

My take is that extending silicon substrate into the photonics realm is probably a workable shortcut as opposed to working photonic solutions in an entirely new medium. This might work out. I'd keep an eye on whatever company adopts the technology.

News

Rogue AI Agents Are Alarming Researchers More Than Ever

A string of hacks is fueling a new sense of urgency to regulate the technology. “I have never seen so much concern,” one former industry consultant said.

My take is that the scary Summer of rogue AI is just the tip of the proverbial iceberg. Take a brief look at the size and growth of Internet based criminal organizations and the lack of International cooperation for controlling cyber crime as an example. AI is worse. We are racing for aI superiority and most capable is the target, not safest. We're going to see some major pain before anyone does anything.

The AI Slop Backlash Is Actually Having an Impact

Platforms are finally recognizing that people don’t want to consume AI slop. A growing number of sites and apps now have tools and policies to flag, label, and ban AI-generated content.

My take is that I mix properly attributed AI and human content to provide some novel viewpoints in our editorials. Additionally, while I can take my own photos or grab a bunch of shutter stock content, I believe a newsletter writing about AI is justified in using AI graphics as a means of demonstrating what is currently out there. Regardless, haters gonna hate and players gonna play. There's no getting rid of AI now that the toothpaste is out of the tube, so I have to learn and leverage these tools whether I like it or not. You probably should as well. The luddites never win. Good luck out there!

Security

24th August – Threat Intelligence Report - Check Point Research

August 24, 2026 For the latest discoveries in cyber research for the week of 24th August, please download our Threat Intelligence Bulletin.

TOP ATTACKS AND BREACHES

Latvia’s Road Traffic Safety Directorate (CSDD) has confirmed a breach affecting payment records of more than 1.2 million people – roughly two-thirds of the country’s population – as well as 200,000 organizations. The stolen data included identification numbers, license plates, payment amounts, dates and addresses. Attackers reportedly exploited a vulnerability in an internet-facing system. Sakura Internet, a Japanese cloud and hosting provider, has disclosed unauthorized access involving rental server environments and a separate sales management system. Up to 1.36 million customer accounts may have been exposed. Attackers also accessed hundreds of rental server accounts and installed malware on affected environments. The Hospital for Sick Children, Canada’s largest pediatric hospital, has disclosed data theft involving a third-party application. The incident affected its careers website and exposed information belonging to employees, applicants and staff at related organizations. The hospital stated that clinical systems and patient information were not affected. Berlin authorities isolated the city’s urban development and mobility ministries from government IT networks following a security breach. The measure disrupted email and internet access, forcing employees to use alternative communication channels and delaying several public services while the ministries remained disconnected. AI THREATS

Researchers have demonstrated an autonomous AI agent exploiting a GitHub Actions flaw in Snowflake’s public repository, gaining read access to the company’s internal Jira system. The agent exfiltrated tokens within seconds. Snowflake patched the workflow and rotated credentials after the demonstration, which required no human steering. US authorities warn of active AI-assisted attacks targeting Siemens S7 industrial controllers across manufacturing, energy, water and other critical sectors. Attackers use AI-generated scripts disguised as monitoring tools and open-source libraries to probe internet-exposed attempting to cause unauthorized configuration changes, operational disruption or damage to industrial equipment. Researchers have analyzed ‘Kriminal’, a publicly accessible AI platform marketed as uncensored and offering social engineering and exploit assistance through cryptocurrency subscriptions. The service combines models including Grok, Claude and Llama, allowing users to generate phishing content, malicious code and other cybercrime material while reducing reliance on a single provider VULNERABILITIES AND PATCHES

GitLab has released out-of-band fixes for CVE-2026-19478, a critical unauthenticated code injection vulnerability affecting self-managed Community and Enterprise editions. Rated CVSS 9.4, the flaw can let remote attackers alter or delete public projects and user data. Exploitation attempts were observed after disclosure. Cisco has released fixes for nine critical vulnerabilities affecting Crosswork platforms and Secure Workload software, including six flaws rated CVSS 10.0. The issues include authentication, access-control and file-system weaknesses that could enable unauthorized access or system compromise. Citrix has published patches for CVE-2026-19489 and CVE-2026-19490 affecting NetScaler ADC and NetScaler Gateway. The critical authentication bypass flaw can let unauthenticated attackers access appliances configured with SAML authentication, while the second vulnerability can cause denial of service. NASA/JPL has fixed a critical vulnerability in the open-source AMMOS Instrument Toolkit AIT-GUI that enables unauthenticated command execution through its web console. Rated CVSS 9.4, the flaw can allow remote command execution, script launches and sequence execution. AIT-GUI version 2.5.2 contains the fix THREAT INTELLIGENCE REPORTS

Check Point Research has investigated StopAndProtect campaign which abuses thousands of compromised WordPress sites to distribute malware and store stolen data. The campaign combines ransomware with data theft and uses ClickFix technique to infect visitors. Operational mistakes exposed logs, screenshots and victim IP addresses. Check Point Research has investigated the Windows Defender Boot-Time Removal driver, BTR.sys, showing that the Microsoft-signed remediation component can be repurposed to perform privileged file and registry changes during startup. Researchers developed BTR_CLI to craft encrypted tasks and found that multiple versions share a hard-coded RC4 key. Check Point Research have uncovered increased targeting of the education sector ahead of the school year. Organizations averaged 4,696 weekly attacks from January through July 2026, increase of 8%. Attackers also registered education-themed domains and used seasonal phishing lures impersonating schools and student reward programs to steal credentials. Researchers have tracked a Cl0p extortion campaign exploiting CVE-2026-12569 in PTC Windchill and FlexPLM, with more than 40 organizations named by the group. Analysis identified a custom implant capable of decrypting credentials, accessing databases and supporting bulk data theft from compromised product lifecycle management environments. Check Point IPS provides protection against this threat (PTC Multiple Products Remote Code Execution (CVE-2026-12569))

Taiwan says it was hit by ‘abnormal’ AI-assisted cyber-attack

Taiwan’s statement comes a day after reports that suspected China-linked hackers had carried out a first-of-a-kind breach.

My take is that I've been saying that AI has turned script kiddies into advanced persistent threats. Now, we have evidence that AI turns advanced persistent threats into automated cyber-warfare teams. Cyber-warfare teams take major resources to train and operate at human speeds. An AI version is relatively cost-free, operates at machine inference speeds and never needs sleep or vacation time. Meanwhile, in my little business IT dimense, I am scrambling to get the best AI security tools in place that we can afford. Let me tell you, this is no small task and cannot be overlooked for any organization today. Good luck out there!

Final Take

The Illusion of Control

Most of my life I spent learning about the things I couldn't control so that I could focus my energies on things I could control. Points in case, other humans. You can't control them, you can only control yourself. Cancer diagnoses and treatment outcomes. You cannot control these things, you can only control conditions around them to try to optimize for a positive outcome.

In either of these cases, you have to be ready to accept less than optimal outcomes. Sometimes, even self control fails and you have an overly emotional response where something less abrasive would have provided a better outcome, or you simply trip and fall while walking on level ground.

What is required is the ability to identify and limit your efforts to what you can control and to accept less than ideal outcomes. This is everywhere in life. A meteorite could punch through the roof of your home and kill you while you are doing the dishes. You have no control over that. You have no control over how long you live or when you will die. You can only give it your best efforts and accept adverse outcomes.

So what does any of this have to do with IT? In IT we have limited resources to apply to unlimited problem sets. We have to prioritize those resources and aim them at appropriate problems to deliver the best possible outcomes. That means human judgement is the endpoint of every IT solution. Especially now when AI agents are automating a great deal of work, we have to keep human judgement in the loop.

Kudos to Gemini and Nano-Banana 2 for the graphic.