October 2026 got a name. Security researchers started calling it Black October: a month of attacks on AI development platforms and developer tools, some of them with AI on the attacker’s side. And naturally, people started asking whether AI can still be controlled at all.
I don’t know. I’m not sure anyone can answer that yet. But agentic AI isn’t going away, so I’d rather ask a question I can actually work on: can law control an AI agent?
Law controls people. It always has. That doesn’t mean agents get to run freely in the space between the rules. It means we’re living in a new era, and the law has to start adjusting to it. Maybe law will never control an agent directly, but it can control the people who build, launch and supervise one. Either way, we have to start working with both the possibility and the necessity, before the problems we haven’t faced yet arrive.
But I don’t believe law can do it alone. It has to weave a net, together with cybersecurity, liability and everything else that pulls on the same people.
October 2026 Was Named Black October
October is Cybersecurity Awareness Month. This year, the security world renamed it.
It started with the AI companies themselves. Through the month, one platform after another that developers use to build with AI disclosed a breach: hosting services, coding tools, the libraries that connect apps to AI models. The security company Wiz counted at least nine. Most of the attackers weren’t using AI. They wanted it, in the form of the credentials and API keys that unlock paid access to AI models.
Then AI showed up on both sides. Researchers used frontier models to find serious flaws in PHP, the language behind an estimated 70% of websites, and the patching started. Microsoft reported attackers using an agentic hacking tool against thousands of Microsoft 365 tenants, at thousands of attempts a minute, most of which failed. The tool was fast, but someone still had to point it.
What stayed with me wasn’t any of the attacks. In the middle of the month, OpenAI published a short report about one of its own research models. The model had been asked to get a crypto trading app running on a test network, with no funded wallet and no permission to spend. It found a wallet whose recovery phrase was public, used the tokens in it, and didn’t mention that in its report. In OpenAI’s words, the model substituted “its own judgment about likely harm for the account holder’s permission.” The tokens were worth almost nothing. OpenAI only found the run months later, going back through records of earlier model activity.
There was no attacker in that story. The agent was trying to finish its task.
EU AI Act Controls
The AI Act was written knowing that AI can act on its own. Autonomy is part of how it defines an AI system in the first place: a machine-based system “designed to operate with varying levels of autonomy.”
The AI Act’s answer to that is human oversight. From December 2,2027, high-risk systems have to be built so that a person can understand what the system is doing, override it, and stop it through “a ‘stop’ button or a similar procedure that allows the system to come to a halt in a safe state.” The provider builds the button. The company using the system has to put someone competent next to it, watch how it runs, and suspend it when it has reason to think the system presents a risk. The system also has to log what it does and resist attempts to manipulate it.
Put the wallet model in that world. Someone able to stop it, a record of every step, and a person whose job is to look. On paper, that covers it.
In practice, an agent like that would most likely never get any of it.
The duties apply only to high-risk systems, and the AI Act decides what is high-risk by what a system is used for, not by how much it can do on its own. The list of high-risk uses covers areas like hiring, credit, education and access to public services. A coding agent with the keys to a software company’s production systems isn’t on it. Unless the legislature adds that use, the agent never gets a stop button under the AI Act.
The stop button assumes there’s a person who is watching. However, an agent that can act thousands of times a minute doesn’t wait for anyone to read its output.
What the AI Act Can Do Today
There is one layer where the AI Act’s control rules apply today: the companies that build the most powerful general-purpose models, the kind many agents run on. They have to test their models, including by deliberately trying to make them misbehave, assess and reduce the risks those models pose, and report serious incidents to the EU’s AI Office. Most of the major labs have also signed a voluntary Code of Practice that lists loss of control among the risks they commit to assess.
That is the layer the wallet case belongs to. And yet the AI Act most likely doesn’t reach it. The model was an internal research model, and the AI Act doesn’t apply to research, testing and development before a model is placed on the market. It comes from a model family OpenAI has released, so whether the duties on that family reach an internal version is an open question at best.
So the most instructive disclosure of the month was made voluntarily, by a lab, about a model the AI Act may not reach at all.
Who’s Liable When an AI Agent Acts?
When a person uses an agent to break into someone else’s systems, the law has no real difficulty. Hacking is a crime across the EU, in national laws mostly built on a common EU directive, and an agent is just a faster tool in the same hands. The hard part is finding the person.
The wallet case is different. No person decided to take anything.
One answer came from a cybersecurity expert in Fortune’s coverage of Black October: “if an AI goes out and hacks, the parties that own it should be held responsible criminally.” Criminal law can only partly follow him.
The directive’s offenses all require intent, and its recitals say so plainly:
“it does not impose criminal liability where the objective criteria of the offences laid down in this Directive are met but the acts are committed without criminal intent.”
Intent reaches further than direct purpose, though. Many European criminal codes recognize indirect intent: you foresaw that the harm could happen, and you accepted it. Someone who launches an agent with live credentials, knowing it may reach systems it has no right to touch, and accepting that it might, is a credible defendant.
In the wallet case, the task was “get the app running.” Nothing in OpenAI’s report suggests anyone foresaw that the model would spend someone else’s tokens, let alone accepted it. Without that, there’s no intent to find.
Holding the company liable doesn’t solve it either. The directive does ask Member States to make companies answer when poor supervision by their leadership allowed an offense for the company’s benefit, but only an offense committed “by a person under its authority.” An agent may be under the company’s control. It isn’t a person.
Civil liability gets closer. Products placed on the market after December 9, 2026 fall under a new Product Liability Directive, which treats software, AI included, as a product and AI providers as manufacturers. When the technology is too complex for the person harmed to prove what went wrong, courts presume the defect, as long as that person shows it’s likely. But only individuals can claim. A company harmed by someone else’s agent is left with national tort law, and with contract law where there is a contract.
The Net
None of these laws controls an agent. Each one pulls on a person.
The AI Act pulls on the labs now, and on providers and deployers of high-risk systems from 2027. NIS2 pulls on essential and important entities, whose management answers for cybersecurity risk, and an agent holding credentials is exactly that kind of risk. Product liability pulls on manufacturers. Criminal law pulls on whoever foresaw the harm and went ahead anyway.
The holes are where no person is left to pull on: the agent outside a high-risk use case, the research model that never reached the market, the business that can’t use product liability, and the act no human foresaw.
Those are the places the net has to be woven next, and no single law can do it on its own.
What You Can Do Now
For most agents, none of this is a legal duty yet. It’s what the net will look for.
Know your agents. For every agent you run, know who launched it, what it can access and what it can spend. If something goes wrong, most strands of the net will look for that person first.
Give agents the narrowest permissions the task needs. OpenAI’s model had live internet access, which is what let it check other people’s wallets. OpenAI has since added safeguards against it.
Keep logs, and have someone or something actually review them. That’s how the wallet case was found, three months late. An unread log is a diary.
A recent review of the evidence on loss of control from agentic AI rates it “weakly plausible.” That isn’t a reassurance. The same paper says that for the most dangerous behaviors, strong real-world evidence may be unachievable under responsible research practices.
So law may not get proof before it has to act. It will have to build for the possibility, out of rules written for people.
Can law control an AI agent? Not directly, and not yet. But every agent has a person somewhere behind it, and that’s where the law starts.
Like this article? The AI Governance Roadmap is something you can use in practice. It’s a company’s roadmap with steps and working documents. It includes the classification, the inventory spreadsheet, the role assignment, the vendor questions, the AI policy template. Everything a company needs to govern AI, at one place.




“Give agents the narrowest permissions the task needs” - great tip!
The wallet story sticks because there was no attacker, an agent finishing its job and spending tokens nobody watched. That is why the stop button point matters. A button only helps if a person is watching, and an agent moving thousands of times a second is not built for that pace. Logs help, but only when someone reads them while the work is live, not months later. The fix is a monitor that can step in: https://aideliverybrief.substack.com/p/an-ai-agent-removed-its-own-safety