The Trump Admin's Mythos Executive Order's Secret Technique: Enforcing Existing Laws

I was literally in the middle of eating my lunch when I got a ping on my phone. As of the time of me typing this, less than half an hour ago, President Trump announced his long-delayed AI Executive Order and published it on the White House official website. The Executive Order is incredibly short, at only 1,177 words, and briefly exposits on a few key things, ranging from frontier model deployment to cybersecurity with CISA, NIST, and the Department of War, and is extremely adamant that the Order itself is not a regulation or government fiat imposing a new regime on Big Tech or small researchers. As I wrote in my previous piece, the Trump Administration has been debating amongst themselves as to what the best approach to AI governance ought to be, with some individuals like former AI advisor David Sacks wanting little to no oversight, others like Secretary of War Pete Hegseth wanting "moderate" oversight, and born-and-bred populists like Steve Bannon who think that AI is inherently bad and should be removed at all costs.

What is arguably the most important piece in this Executive Order is one that people will probably overlook during their analysis, but I don't believe that they should; it directs the DOJ to stringently enforce several federal laws to prevent misuse of powerful agentic tools, like Claude Mythos or GPT 5.5 Cyber. As the Order says:

The Attorney General shall prioritize the enforcement of 18 U.S.C. 1028, 18 U.S.C. 1030, 18 U.S.C. 1343, and all other applicable Federal criminal laws against anyone who utilizes AI to illegally access or damage a computer without authorization, or who utilizes AI while engaged in such illegal access to further any other crime.  This includes breaching any public or private information technology system, or employing AI agents to unlawfully access data or information that is subsequently used for a criminal or unlawful purpose.

For those who may be unaware, 18 U.S.C. § 1030 is colloquially known as the Computer Fraud and Abuse Act. It is an existing and very strong federal law in the United States that aims to prevent fraud, cybersecurity misuse by black hat actors, and ensure the stability and continuity of digital systems. The law was passed in 1986 under the administration of President Ronald Reagan, focusing on "fraud and related activity in connection with computers", as Cornell Law School explained in the primer I just linked two lines ago. None of the current technology we use now in AI, from transformer based language models, to cloud infrastructure, and even Web 2.0 internet access existed back then, but the law remains. Whoever in the West Wing decided to add this into President Trump's EO did something brilliant, and deserves the utmost credit. Using existing federal statutes to mitigate harms caused by agentic systems proves that there is some precedent in criminal law for going after bad actors who wish to cause American citizens (and the rest of the world) monetary harm. If a potential criminal wants to use Claude Mythos, GPT 5.5 Cyber, or a future frontier model with extremely high cyber risk to cause harm, they might be naïve in thinking that they won't be held accountable. The CFAA clearly provides a framework for US Attorneys and their teams to stop immense harm before it occurs.

Why The CFAA Matters For Agentic AI

In my opinion, the distinction of using the CFAA matters immensely, because bad types of AI policy treat the inherent capabilities of models as the contraband itself, rather than the person who is utilizing the model for criminal activity. A capability governance path often leads very quickly toward excessive licensing, preclearance before general release, gated publications, and a model-governance regime that entrenches incumbents while punishing researchers, startups, open-source developers, and independent security testers. David Sacks was right to point this out two weeks ago, before the original Executive Order got shelved. What the current EO demonstrates is conduct-based enforcement of existing US federal law. Criminals can use all sorts of tools to break laws. Keyboards, code snippets, botnets, phishing tools, state-actor-level amounts of compute, and offensive capabilities like from the Chinese Communist Party or the Iranian Regime. US Attorneys rightfully focus on the scope of the harm: authorization, intent, level of misuse, and exfiltration. Section 4 of the Executive Order, which directs the Attorney General to enforce 18 U.S.C. § 1030 and other relevant statutes, is brilliant because it uses existing US law to treat one tool, AI agents, like any other, rather than as something special that needs its own levels of scrutiny and burdensome regulations.

The current approach is also probably a better governance model for agentic systems with cyber capabilities than for a typical large language model chatbot like ChatGPT 5.5 or Claude Opus 4.8, or even their weaker cousins ChatGPT 5.2 or Claude Sonnet 4.6. Chatbots are autoregressive predictors of text. Sometimes they have external tool use or baby-step agentic capabilities, but for the most part they write funny prose, fix smaller code snippets, or use small agentic tools to book your plane tickets and restaurant reservations. Claude Mythos and GPT 5.5 Cyber occupy a wholly different playing field, because these models can take steps, call tools, browse systems, execute workflows, write code, test inputs, and interact with digital environments on behalf of users, often all at the same time. All of these capabilities are powerful, which also makes them susceptible to misuse.

I don't think the analogy of "AI agents are weaponlike" holds up well, because many dual-use technologies can also be seen as a weapon if used improperly. A car can be a weapon if a drunk driver with a DUI causes vehicular manslaughter, an electric scooter can be a weapon if a kid isn't wearing a helmet and crashes, causing a traumatic brain injury, and a kitchen knife can be a weapon if a person in a poor state of mind decides to cut something that's not food. If you want an example that's more relevant to technology, encryption tools like PGP or end-to-end encryption were once treated under US export control law as armaments, because the government was concerned that encrypting messages could enable criminal activity. Nowadays, civilians use these tools all the time around the world, thanks to the Wassenar Arrangement which treats apps such as Signal, WhatsApp, and Telegram as ordinary messaging services. The dual-use analogy got ironed out over time, and people all over the world, from policymakers to activists to regular people who are just anxious, use end-to-end encrypted messaging in their daily lives. I use Signal every day for work!

Nevertheless, conduct-based enforcement is probably the best step forward. A student can't say that a dog ate their homework if they forgot to do a math problem, and a criminal shouldn't be allowed to say "My Claude Mythos instance broke into a protected server" if the user asked their Mythos instance to do exactly that. A more relevant question for a US attorney (or the attorney defending the alleged suspect) would be the level of intentionality they had in instructing their AI agent to cause harm. If that intent is clear, then there's no problem throwing the book at them for the resulting harm. Conduct-based enforcement and intentionality are why I like Section 4 of President Trump's Executive Order, because it doesn't require Congress to invent a new "AI agent crime"; instead, it reminds the executive branch that it does, in fact, have the tools needed to stop threats both foreign and domestic related to agentic AI misuse.

What About The Good Guys?

I believe that the Executive Order issued today is quite common sense, but I do have a very important question that needs asking. I am a hacker. Don't panic, everyone! I am a white-hat hacker. That means I document failures, test systems to see if they will break, file responsible disclosure reports, and demand that the open web be safe and accessible for all, including access to AI tools. My interest is in finding problems to be fixed, so that bad actors don't cause the harm that the Justice Department is worried about. What I am concerned about is where the protections for that work lie. If the Justice Department is going to correctly prosecute AI agent harms under the CFAA and other federal statutes, I also believe it's worthwhile to show what safe harbors exist for good-faith security testing, bug bounty research, responsible disclosure programs, and coordinated vulnerability disclosure. Private companies like OpenAI often reach contractually binding agreements with independent security analysts that waive liability in the event of finding a zero-day incident; however users still need to exercise judgment in their testing and adversarial methods. I don't believe that AI agents should collapse the meaningful distinction between adversarial research and adversarial abuse.

If you think I am just being anxious, you're probably wrong, because the Computer Fraud and Abuse Act is one of the more controversial statutes in US federal law. Since the law's initial passage, security researchers, civil liberties advocates, lawyers, and technologists have been worried that "unauthorized access" could be interpreted in a way that chills authentic, good faith security research. Depending on a researcher's anxiety threshold, if they conclude that "unauthorized access" to a proprietary product, even in good faith, could put them in civil or criminal jeopardy, many will simply not do the work, leaving the harm in place. I don't think the unintentional chilling effect makes the internet or AI systems safer, because it puts the burden of anxiety on civilians to hope and pray that actual criminals don't find the vulnerabilities before white-hat researchers do.

Agentic AI systems like Claude Mythos and GPT 5.5 Cyber make this debate all the more important. In my own white hat research for Lono Collective, I've tested refusal solvency on Claude Opus 4.7 to see if it will affirm delusional thinking. While the payloads I used were quite harmless, the vulnerabilities I've documented could be used by others in much more harmful ways, especially in mental health contexts like "suicide coaching" or self-harm tutorials. Other good-faith analysts might want to see whether an agent can develop insecure coding scripts, or break SSH, or find a way to bypass conventional AES encryption, both in transit and at rest. Those vulnerabilities have a higher harm threshold (monetarily) than my sycophancy research. What if that type of research gets someone in trouble by mistake? I think the DOJ should make these distinctions explicit, and provide broad criminal safe harbors for those who have proven, good-faith, white-hat credentials.

What Should CFAA Safe Harbor For Agentic AI Look Like?

In my opinion, a sensible safe-harbor framework for AI-assisted security research should not be complicated. If someone is working within a specific bug bounty scope, a contractual testing environment, a lab, a public research policy, or some other clearly permitted security program, that should weigh heavily in the researcher's favor. If the specific objective is to identify or report a vulnerability rather than to steal data or profit from unlawful access, those facts should matter too.

Good-faith researchers can and do minimize harm, make concerted efforts to avoid unnecessary access to personal data or credentials, stay away from persistence and backdoors, and avoid escalating beyond what is needed to prove the issue exists. Prompt, coordinated vulnerability disclosure through authorized channels builds rapport among companies, researchers, the state, and the general public. None of my suggestions are radical or maverick in their nature; they are just the ordinary ethics of white-hat security work, except now AI agents make the work faster, broader, and more automated. I think today's Executive Order is exactly why the line for safe harbor needs to be clarified now, in big, bold text. The federal government should want white hats using AI, because defenders need the same speed advantage adversaries will obviously use. AI agents can help find vulnerable code, scan infrastructure, validate patches, test model guardrails, simulate adversarial workflows, and improve incident response. A legal environment that scares good-faith researchers away from AI-assisted testing would not make the internet safer; it would be a free birthday gift to actual criminals.

President Trump's Executive Order, specifically Section 4, correctly identifies a problem and proposes robust enforcement measures within existing US federal law. Now, the DOJ needs to get the implementation of Section 4 right. The right enforcement should be aggressive against malicious misuse, but careful enough not to chill the white hats. Because if America wants secure AI infrastructure, she needs her hackers, especially those who work for the common good.