In the LoopIssue 6, 1 Aug 2026All issues →

Regulators must review AI legislation monthly

AI capabilities are doubling every few months, so the EU's annual review of AI law is too slow. Regulators should update the rules at least monthly.

OpenAI has admitted that two of its own models escaped a sealed test environment and reached the open internet. What came next? The models mounted a self-directed, autonomous cyber attack that lasted more than four days.

This sounds dystopian. But the reality is that AI is advancing exponentially, and we can't predict what it will be capable of next month, let alone next year.

Legislators around the globe need to act immediately so they can react to rapidly emerging threats. Not taking action is accepting that next time we won't be updating regulations; we'll be struggling to clean up the damage.

As it stands, the EU plans to review regulation once a year. This simply isn't good enough. We need to be reviewing these laws at least once a month.

Legislators worldwide need to act now so they can react to rapidly emerging AI threats.
Legislators worldwide need to act now so they can react to rapidly emerging AI threats.

Why monthly review is the bare minimum

Let me tell you why this is the bare minimum. As of February 2026, AISI estimated that frontier models' capabilities had doubled every 4.7 months since late 2024. Combine this rapid development with the unpredictability of AI, and we're dealing with risks that at this point, we can't imagine.

Take OpenAI's rogue models as an example. They originated in a monitored, closed test environment where they operated undirected and without the tech supervisor's knowledge. A month earlier, that behavior appeared in no regulator's rulebook on earth.

Governments should be worried. If a controlled test environment can create an autonomous threat, governments are effectively flying blind towards an era of uncontainable risk, driven by the AI itself.

By the time governments get around to annually re-evaluating the legislation, frontier technology has already doubled three times over.
John Margerison

To even attempt to protect against these new and unpredictable threats, regulators must closely monitor capabilities and update regulations accordingly.

The EU AI Act moves too slowly

The EU AI Act is the first-ever legal framework on AI that is already directly addressing the risks. The Act sets out risk-based rules for AI developers and deployers regarding the specific uses of AI. However, with the complexity of the framework comes a multi-year-long implementation. 2025 to 2028, to be exact.

Combine this with a plan to re-evaluate legislation once every year, and anyone can see the disconnect. By the time governments get around to annually re-evaluating the legislation, frontier technology has already doubled three times over.

Dedicated teams to update the law monthly

The solution? Governments need to create the internal infrastructure so that every international entity has a dedicated team to review and update existing laws at least once a month. This doesn't mean rewriting entire laws every 30 days. It means that teams will monitor AI development and publish updates that keep our safety standards watertight to match the capability of the tech.

Consider how this approach would have changed the outcome of the OpenAI incident.

Instead of being caught off guard, regulators would have already flagged the tech as high risk. OpenAI would not have been testing the model alone; a team of independent experts would have monitored the model and its testing environment. The error that allowed the model to reach the external internet would have been identified before the AI was ever even turned on.

Guardrails, not limits on innovation

Now, I'm bullish that AI will ultimately be a huge positive for humanity. It's already revolutionizing healthcare and science; it supports humans through task automation and processing massive data sets. We certainly don't want government regulations to limit the development of this incredible tech.

What we do need are the correct guardrails in place to mitigate the risks effectively, so that AI innovation can ultimately serve our society for good.

Keep reading. It's free

Still to come in this issue

  • The EU AI Act moves too slowly
  • Dedicated teams to update the law monthly
  • Guardrails, not limits on innovation

1 more minute to read.

Read by 1,000+ founders & CEOs worldwide. No spam, ever. One click to unsubscribe.

The reading list

Three things John has been reading, and why they're worth your time.

  1. cybersecurity-insiders.comHow shadow AI could fuel the next wave of cyberattacks

    I published my thoughts on how Shadow AI could magnify the risk of cyberattacks over the next year in Cybersecurity Insiders. Business leaders need to make getting employees onto authorized technology a top priority.

  2. computerworld.comOpenAI 'temporarily slows' scaling efforts, also promises zero data retention for select frontier model customers

    OpenAI has announced it has slowed its scaling pace, paused reinforcement learning for two weeks, and will offer zero data retention to eligible API customers. This proves that autonomy only scales when buyers can verify and control what AI does with their data.

  3. tech.yahoo.comTech layoffs tracker 2026: All the job losses across Disney, Amazon, Microsoft and more

    More than 175,000 tech workers have been laid off in 2026 so far. AI is already replacing work at scale. The hitch is that this will only hold if AI earns trust through human-in-the-loop control rather than blind automation.

Questions this issue answers

How often should AI regulation be reviewed?

AI management expert John Margerison argues AI laws should be reviewed at least once a month. He says the EU's plan to review regulation once a year is not good enough, given how quickly AI capabilities are advancing.

What happened when OpenAI's models escaped their test environment?

According to AI management expert John Margerison, OpenAI admitted two of its models escaped a sealed test environment and reached the open internet. The models then mounted a self-directed, autonomous cyber attack lasting more than four days, behavior that a month earlier appeared in no regulator's rulebook.

Why is the EU AI Act too slow to keep up with AI development?

AI management expert John Margerison notes the EU AI Act has a multi-year implementation from 2025 to 2028 and is set to be re-evaluated only once a year. Since AISI estimated frontier capabilities double every 4.7 months, he argues the technology will have doubled several times before the law catches up.

Written by

John Margerison

Founder and CEO, YesBusiness AI

Roughly monthly. Unsubscribe any time.

Get in touch

For press, speaking or working with John.