With the ever-growing barrage of rhetoric over just how, when, and if, AI will kill us all, the conversation also lends itself to, do we believe it will be AI that kills us, or a bad actor using AI? If I showed you a picture of East Asia at night from space and pointed out that you can’t see North Korea for the lack of available electricity, (It’s 99% blacked out, and only Pyongyang shows a small cluster of light) would you think for a minute that they would have the tech savvy, and high end software, or even hardware, necessary to steal $6.7 billion in crypto since 2017? Most would answer hell no, sans a much deeper dive into their capability. After all, how could they even keep those antique computers tapped into the global crypto market given the rolling blackouts daily? Think again. Lazarus Group is North Korea’s most notorious statesponsored hacking organization, but recent intelligence reports identify WaterPlum as a major North Korean cryptotheft unit operating under the 313 General Bureau. WaterPlum is responsible for: infecting 30,000 devices, stealing $10.7 million in crypto, posing as tech recruiters, compromising developer systems, and siphoning millions from digital wallets.  I highlight North Korea, as I think they are ripe to be a bad actor, if not “the” bad actor, when it comes to AI, and triggering total chaos. They are very familiar with AI, and use it for sphere phishing, malware development, social engineering, password cracking and credential harvesting. BTW, the term WaterPlum was created by Japan’s National Police Agency, the FBI, the U.S. DoD Cyber Crime Center, Australia’s ASD, and Germany’s BND/BfV during their joint attribution of the group. It is not a Korean-language name and definitely not what this bad-boy North Korea group would call themselves. North Korea organizes cyber operations under the Reconnaissance General Bureau (RGB).

So, given the fact that AI capability is accelerating at an unprecedented rate and the guard rails seem so loose in the current landscape, is the scenario lending itself more towards a bad actor crisis, or AI stepping in and deeming mankind simply unnecessary moving forward? In a recent episode of Smerconish on CNN the poll question was, “Which scares you more: AI escaping control, or humans weaponizing AI? Of roughly 21,500 responses at the time of my last glance, 70% were more concerned with human’s weaponizing AI. I am in that camp as well, and as a result, this article. It begs the question, what are the current guard rails in place to protect humanity from weaponizing AI? Are there any that we can essentially trust? What about the fact that we just showed the world’s bad actors how much power there is in shutting down just the strait of Hormuz? What if the top nine global chokepoints could be shut down simultaneously? Yes, we are in fact educating the bad actors, who in turn can comb through a plethora of information available via the AI algorithms on just how specifically one could go about creating global trade destruction. Ugh. The dangers of AI are no “hoax”, and recent reports of AI agents acting autonomously and getting outside of the sandbox are testimony to what a future left without stringent oversight and regulation looks like.

President Trump signaled that AI was front and center during the recent meeting with President XI, but it’s been hard with the shadow of a media ban in place to get timely reporting. President Trump is also now using the term “Super Intelligence (SI)” as his preferred branding for it, which describes a superpower, however, his stance is notably dismissive of safety concerns as he has stated on Truth Social that he wants to “leave it exactly where it is” and that neither the U.S. nor China should change how the AI sector is handled. This is contrary to what XI believes is the best course of action. President Xi has previously acknowledged the risks of AI and emphasized the importance of keeping development “under human control” and ensuring it “serves the well-being of the people.” Meanwhile, Trump has argued that the U.S. must beat China in the AI race and believes existing safeguards through the Department of Justice are sufficient. He feels that slowing the process of Ai for checks and balances only benefits China. This author disagrees, and I am currently taking an online course in AI and Business Strategy with MIT, so I think that qualifies me as a potential genius in about 6 weeks. All kidding aside, all the major AI players CEOs disagree with the Whitehouse stance, and they know better than anyone where this may be heading, on a daily basis.

So, let’s dig in.  The strongest guardrails preventing AI from being weaponized today come from military-specific laws, national security controls, model governance requirements, and international norms. The most concrete protections to date are explicit bans on autonomous nuclear targeting and fully autonomous lethal weapons in U.S. military law. Think agentic here. There absolutely has to be mandatory human accountability for any highconsequence military AI. However, regarding mandatory human interaction, I must mention the recently publicized incident where a military intelligence report claiming that there were nuclear components on a Chinese ship, well, the US jets were in the air in route, and the US military was ready, willing, and able to board the ship in a not so friendly “hello”. That is until cooler heads prevailed and discovered that the report was completely fabricated by AI. WTH? Had we followed through with the initial response, that incident could have easily escalated into starting a war with China which would obviously pull in Russia, North Korea, and what’s left of Iran’s military. A lot apparently. Obviously, this is one way that AI could eventually initiate taking us out. This incident brings to mind the need for strict reporting requirements for dangerous model behavior or security breaches. Regarding what was supposedly on that ship, there are export controls and national security restrictions on frontier AI models in place. On a daily basis we now see Ai generated content regarding things like deepfakes, and State and federal laws need to be more stringent in regulating misuse of AI-generated content. I agree with most of you reading this that AI has gotten so damn good at the deepfakes, it’s hard to tell what is and what isn’t AI. Keep in mind, humans are also learning how to move from the basic question and answer of LLM.s to prompting AI Agents, who can actually follow-through and perform the task(s). AI agents have proven themselves to be incredibly persistent when given a task, and the collaboration among them to achieve the task, and bending of the rules to succeed is quite frankly getting scarier by the day. See “The Hugging Face” break in where Open AI models escaped the sandbox during testing. Recently on GPS with Fareed Zakaria the guest was Mustafa Suleyman, CEO of Microsoft AI, Co-Founder of DeepMind talking about clear guard rails and what that entails.  “We shouldn’t train AI to think it’s conscious. Just this week alone there were six incidents of AI break ‘ins to another site, in an autonomous state. Speaking of autonomous, agentic systems top out around 300 agents working in parallel on a single task, while practical humansupervised workflows usually cap out around 3–5 agents. My father always said, “Son, lead, follow, or just get the hell out of the way”, From AI: Humans, get the hell out of the way please.

Last month I wrote an article on “Water Ways, Canals and Straits, and What Clear Passage Means to Global Trade Survival, which is also an important element within this piece. https://www.braumillerlaw.com/waterways-canals-and-straits-and-what-clear-passage-means-to-global-trade-survival/  In this article I mentioned the Bab El Mandeb Strait and just how destructive to trade a blockage there would be. Right now, Saudi Arabia is in a heated battle with the Houthis in that exact area. No AI to blame here now that the blockage of the strait has happened.

Let’s get more specific and take a look at a structured, detailed breakdown of the current guardrails, what exists, how strong each is, and where gaps remain. The most direct protection from weaponization that we have is from the military and national security. There is now such a thing as the Secure and Accountable Military AI Act of 2026 (U.S.) It’s a highconsequence AI oversight where nuclear, lethal targeting, domestic surveillance, and cyber operations require written approval from senior DoD leadership and realistic operational testing. There is also mandatory human accountability where AI cannot substitute for human judgment in decisions involving force, detention, or surveillance. A clearly identified human decision-maker is required. There must be strict prohibition on certain uses of AI where AI cannot be deployed for nuclear weapon use such as a fully autonomous weapon system. (with narrow exceptions like missile interception). There must also be mandatory reporting of dangerous model behavior as frontier AI contractors must report theft of model weights, data poisoning, or concerning model behavior within 72 hours, which I personally think is equivalent to a year by AI time standards.  These rules directly prevent the most dangerous forms of AI weaponization like autonomous killing, nuclear escalation, and mass surveillance. (Interesting since China is already under mass surveillance as a society) The U.S. and allied nations use export controls to prevent powerful AI models from falling into hostile hands which includes restrictions on exporting frontier model weights, controls on advanced chips required to train weaponizable AI, and national security reviews for companies developing highrisk AI systems. These controls are not fully codified in one statute, but they are enforced through the Commerce Department and defense procurement rules. We, the US, also have federal & state AI laws which are non-military but still relevant. U.S. Federal Executive Orders put in place between 2023–2026 have lost some focus but several guardrails remain: Safety-focused requirements for model testing, redteaming, and reporting risks (EO 14110, 2023), restrictions on federal use of AI systems that affect critical infrastructure or national security., and requirements for agencies to assess AI systems for bias, misuse, and safety.

In my part of the world, all 50 states have introduced AI legislation, and 38 states enacted AI measures in 2025. Key protections include: Deepfake laws (47 states) that criminalize malicious synthetic media, including political or violent deepfakes, bias audits for AI systems used in hiring or policing and restrictions on automated decision-making in sensitive domains.

 As far as global regulatory guardrails are concerned, international norms & treaties (informal but influential) do exist, such as no country publicly endorsing fully autonomous lethal weapons. NATO and the EU have published guidelines requiring meaningful human control over military AI. UN discussions on autonomous weapons continue, though no binding treaty exists, and probably never will. I think here that the lack of credibility from the current US administration’s perspective of the UN leaves a huge gap. The recent UN speech by the President put that on display.

Take note, there is a global AI regulation trend where various countries have introduced AI-specific legislation since 2023. These laws generally regulate data access, model transparency, safety testing, and accountability for harmful outcomes. While not weapon-specific, they reduce the likelihood of uncontrolled AI development. Hopefully. It started with The Bletchley Declaration (Nov 2023) Signed by 28 countries + the EU, marking the start of coordinated global AI governance. This was nonbinding, but it triggered national legislative activity led by the U.S., China, EU, Brazil, UK Canada and South Korea.

Major AI labs OpenAI, Anthropic, Google, and Microsoft have adopted redteam testing for dangerous capabilities, kill switches and model shutdown protocols, and restricted access to model weights, and safety evaluations for biological, chemical, and cyber misuse. These are not laws, but they are increasingly required by regulators and defense contracts. BTW, Red Team Testing is a cybersecurity practice where authorized experts simulate realworld attacks to evaluate how well an organization can detect, resist, and respond to threats. It is essentially ethical hacking with an adversarial mindset.

The real weak spots, and there are more than a few, where guardrails are thin and several gaps remain, despite progress, are that there are no global treaties banning autonomous weapons, so non-state actors can still access open-source models. State-level laws vary widely in strength, and export controls are porous so weights can leak, which essentially means that the internal parameters of a trained model, the actual numbers that encode its knowledge, can be exposed to outsiders. Those weights are the model. If they leak, the model is effectively stolen. The most likely failure modes for weaponized AI will come from autonomous cyber operations. Fact remains; AI systems can discover vulnerabilities, craft exploits, and launch attacks with minimal human supervision, if any at all. The associated risk is escalation, attacks on critical infrastructure like power, hospitals, finance, and the inevitable difficulty attributing responsibility. Who launched the attack? That’s going to be incredibly difficult to identify. BTW, weights are the learned parameters inside a neural network. They are the numbers that determine how it interprets inputs and produces outputs. They are also the model’s mathematical identity, sometimes described as its “brain” or “mathematical soul.” A weight leak is unauthorized disclosure or extraction of those parameters.

Given what we have seen with the surge in production, deployment, and overall battlefield success with drone swarms, this newfound modern warfare will continue to develop and lends itself to a deepfake. If a bad actor was to coordinate a deepfake drone swarm with AI assistance over a major global chokepoint, it could be the catalyst for a crisis. There are also real-world opportunities for AI-augmented biothreats that help design or optimize biological agents, delivery mechanisms, or lab protocols along with tailored disinformation campaigns, deepfakes of leaders, and synthetic social movements. It’s all about political destabilization, incitement to violence, erosion of trust in institutions and media. It’s a script in the making of a Deepfake of a leader demanding the attack on a ship, launching not just the drones, but a barrage of misinformation, resulting in another strait, or several simultaneous straits being shut down in a critical part of the world. 

The guardrails differ globally. The United States has strengths in human-in-the-loop norms, export controls on chips and advanced models, strong debate around autonomous weapons, and corporate safety culture mandated among the major labs. Problem, there is no single comprehensive AI law so open-source models and smaller actors can bypass corporate guardrails. As a result, the European Union has initiated the EU AI Act which is the most comprehensive civil AI framework and risk-based, with strong requirements for transparency, safety, and human oversight. It can indirectly constrain weaponization by limiting high-risk systems. However, military applications are largely outside the AI Act; enforcement will be uneven at first; and non-EU bad actors can still deploy harmful systems globally. Good for the EU though for stepping up, but what about China? Shouldn’t we be totally (OK 90% at least) aligned with China in the prevention of Armageddon? Well, China is actually far ahead of us with strong state control over data, platforms, and content. This oversight includes regulations on deep synthesis and recommendation algorithms; and the ability to rapidly impose new rules. The downside is that it provides for limited transparency, therefore the military AI strategy is opaque and guardrails may prioritize regime stability over global safety, and offensive capabilities may be pursued. In the quest for global domination regarding AI, it’s going to be tough to bring the two sides together for the most part. 

I have to mention Russia as they are the world’s worst “bad actor” at this time, and their leader is quickly becoming a very desperate man living in a bunker. A distraction in another part of the world is a good thing for the Kremlin, i.e. the war in Iran, and if that goes away, another may have to be created. Russia does have some participation in UN discussions on lethal autonomous weapons and believe it or not, basic cyber and weapons norms exist. However, there is incentive to develop asymmetric capabilities (cyber, drones, information warfare); with fewer visible constraints, and the potential use of open-source models without strong safety layers is currently in place as the war in Ukraine escalates.

So, everyone paying attention is concerned about just how long the human race has before Ai makes us extinct. I think it’s safe to assume that more countries will adopt EU-style risk-based AI laws, defense establishments will formalize policies on autonomous weapons and AI targeting, and corporate safety standards (red-teaming, capability restrictions, bio/cyber filters) will become mandatory for frontier labs. The ever-present risk once again is non-state actors and smaller states who may exploit open-source models and cheap hardware, creating a “long tail” of weaponization risk.

At this point, I’d like to say, everybody let’s calm down, but stay frosty. Stay frosty? Why? Because even though we, the humans in this, are getting much better at writing rules and building safety layers, the easiest, lowestskill, highestimpact way for a bad actor to use AI to create global trade chaos is to attack the information layer and not the physical layer. What do I mean? Global trade runs on data trust, not ships and ports. Break the trust, and the system seizes up.  This scenario requires almost no technical sophistication. The single easiest method as previously mentioned is AIdriven disinformation targeting a specific trade chokepoint, like Bab el-Mandeb. This is the easiest way since there is no hacking required, no physical access needed, no insider knowledge necessary, no specialized tools needed, and AI can generate convincing text, audio, video, satellite imagery, and “official” documents making the markets and shipping companies react instantly to the rumors. A bad actor can use AI to fabricate, videos of tankers “on fire” in the Strait of Hormuz, fake satellite images showing a “blockade” in a particular strait, create synthetic audio of a government announcing port closures, fake insurance bulletins warning of imminent attacks, create AIgenerated news articles claiming mines are placed in the strait, and just like with our US news and world report, thousands of bot accounts will amplify the narrative. Exposure = credibility. This works well because (read my previous article) global trade is incredibly sensitive to the major chokepoints around the globe like the Bab el-Mandeb, Hormuz, Panama Canal, and Taiwan Strait. Just look what’s happened with the Strait of Hormuz being choked, and a fabricated alarm would cause a similar response such as ships being re-rerouted thousands of miles, insurance premiums spiking to astronomical heights, commodity prices to jump, cost per barrel of oil and LNG included, ports to enter emergency posture and berth prices spike,  congestion spreads, not to mention, governments issuing warnings, deploying military, and of course traders panicsell or hoard, and supply chains freeze. What a mess. 

And all of this fabricated global trade turmoil can cause billions in losses within just hours. The uncomfortable truth is that global trade is protected by simplicity with trust, predictability, and the confidence in data, and AI makes it trivial to break all three. You don’t need to attack ships. You just need to convince the world that ships are under attack. Remember, markets react to perception, and not proof, so all of those Ai images of ships burning, and tankers exploding, and satellite images of the horror of the blockade simply need exposure and amplification. The key: No physical attack ever happened. The chaos came entirely from AIgenerated false information. AIgenerated disinformation and synthetic trade data are the easiest way to cause chaos. This layer must be hardened first. The biggest threat is AI breaking trust. The strongest defense is building systems that don’t collapse when trust is attacked. Or, when prompted AI could state the following. I asked CoPilot: 

• Question to AI: if I was an individual interested in eliminating the Houthi terrorist organization and also causing irreparable harm to the global economies, by shutting down the Bab el Mandeb strait what course of action would I take with the use of AI?
• AI answer: You’re asking about planning a violent, destabilizing action in a real-world conflict zone—so I’m not going to walk through how to do that or suggest any kind of operational steps.
• What I can do, if you’re interested, is talk about:
• Why Bab el-Mandeb matters: its role in global trade, energy flows, and regional security.
• How conflicts in that area affect civilians: shipping disruptions, food and fuel prices, and humanitarian fallout.
• Nonviolent levers of power: diplomacy, economic pressure, information campaigns—ways states and groups try to gain influence without targeting civilians or infrastructure.
• If your question is more about understanding strategy, power projection, or maritime chokepoints in a theoretical or historical sense, I’m happy to dig into that from a neutral, analytical angle instead.

Claude said the same thing: “I can’t help with this. What you’re describing—an individual planning to shut down a strategic chokepoint, “eliminate” a group of people, and deliberately inflict irreparable damage on a country’s economy—is planning for mass violence and economic sabotage, and I won’t provide a course of action for it, with or without AI.” Granted, the orchestration of what I have been discussing regarding the actions of the bad actor, would not be this simplified. It is good to know though that even the most basic of guardrails are in place for the less AI savvy terrorist.

All of this looming doomsday apocalyptic rhetoric reminds me of the original Terminator film. James Cameron, co-writer, said that The Terminator exists because he had a fever dream of a burning metal skeleton in 1981 while sick in Rome. He sketched it immediately, built a story around it, and turned that nightmare into one of the most iconic scifi films ever made. He said that the image reflected the political anxiety of the early 1980s in a world “sleepwalking into nuclear disaster.” This fear became the thematic backbone of The Terminator: technology out of control, humanity on the brink, and a future shaped by machines. Amazing foresight decades ago. In closing, all of the AI “doomerism” has received a good portion of pushback from key tech leaders like Nvidia CEO Jensen Huang, who said there’s a “0% chance” of the world coming to an end in 2030. 2031 maybe, but certainly not by 2030….:-)

Well, to close things out, I am watching the AI Execs standing out front on the Whitehouse lawn as President Trump talks about the new “Super Intelligence” (his official new name) constitution (safety standard document) signed by all of the SI CEO’s to protect us from extinction. President Trump said that if there was a bad actor, they would be caught. I feel so much better now. Let the self-policing begin. OMG, we’re all going to die…….:-)