The EU AI Act, Explained: What Changed in July 2026, and What It Means for Your AI

What the EU AI Act actually is

It is a safety law for AI, built on a simple idea: the more harm an AI system could cause, the more you have to do before you release it.

Every AI system falls into one or more, of four buckets.

Banned. A short list of uses Europe has decided are simply not allowed. Social scoring of citizens. Systems that manipulate vulnerable people. Guessing someone's emotions in a workplace or a classroom. There is no paperwork that makes these acceptable. You either avoid them or you break the law.

High-risk. AI used in decisions that seriously affect someone's life. Screening job applicants. Deciding who gets a loan. Grading students. Medical devices. These carry the heaviest requirements: testing, documentation, human oversight, monitoring after launch.

Transparency required. AI that talks to people or creates content. Your customer service chatbot lives here. So does anything generating images, video, audio or text. The requirement is honesty, split between two roles: if you build the system, tell people they are dealing with AI and embed machine-readable marking in what it generates; if you use it, put a visible label on deep fakes you publish. A system can be high-risk and subject to transparency rules at the same time (a CV-screening tool that also chats with candidates may be both high-risk and subject to transparency rules).

Everything else. Spam filters, recommendation engines, most internal tools. No specific obligations.

One thing worth getting right

A common misunderstanding, including in earlier versions of this article, is that any customer-facing AI chatbot counts as high-risk. It does not.

"High-risk" is not a judgment call about whether your AI feels important, it is a specific published list of uses. A support chatbot is not on that list, but a CV-screening tool is, because hiring is; a credit-scoring model is too, and a tutoring product used in formal education might be.

This matters practically. Teams that assume they are high-risk build expensive programmes they do not owe, and often miss the simpler rule that actually applies to them. Teams that assume they are not high-risk when they are, get caught out badly.

What changed in July 2026

The original timeline assumed a support system that never actually arrived: the technical standards were running late, the official guidance was still in draft, and many EU countries had not even appointed the regulators meant to enforce the law. Rather than let the rules take effect before anyone could realistically follow them, the EU passed an amendment called the Digital Omnibus, which became law on 27 July 2026, six days before the deadline it moved.

What got pushed back: the heavy requirements for high-risk AI. They now apply from December 2027, or August 2028 for AI built into regulated physical products like medical devices and machinery. Separately, generative systems already on the market before August 2026 have until 2 December 2026 to implement machine-readable marking of their output.

What got softened: the AI literacy duty. Article 4 used to require you to ensure your people had a sufficient level of AI literacy. It now requires you to take measures to support it – an obligation of effort rather than result. It still binds every provider and deployer, and national supervision of it starts 3 August 2026.

What did not move:

Since February 2025 - The list of banned uses.

Since August 2025 - Rules for companies that build general-purpose AI models.

August 2026 - Transparency rules. Tell people they are talking to AI, and label AI-generated content.

August 2026 - Regulators can now actually fine model builders. The rules existed before this; the enforcement powers did not.

What is genuinely new:

From 2 December 2026, two further practices are prohibited outright: AI that generates or manipulates intimate images of identifiable people without their consent, and AI that generates child sexual abuse material. These did not exist in the original text – the Omnibus added them.

It does not just ban tools built for that purpose. It also catches ordinary image, video and audio generators where this output is reasonably foreseeable. and reproducible result that does not require significant technical modification to obtain, and the provider has no reasonable and adequate safeguards to reliably prevent it – and to correct misuse once it is observed or reported.

You need safeguards that actually hold, and you need to be able to show they do.

The fines are serious. Banned uses reach EUR 35 million or 7% of worldwide annual revenue, whichever is higher. Most other breaches reach EUR 15 million or 3%.

What this means for you

The most useful way to think about the AI Act is this: it asks you to prove things work, not to promise that they will.

Look at what it actually requires:

- Testing that is documented

- Risk management that is ongoing rather than a one-time review

- Systems that hold up when someone deliberately tries to break them

- Monitoring after launch

- Safeguards that reliably prevent bad outputs.

None of those are things you can write down once and file away. They are things you have to keep demonstrating, especially because your AI keeps changing. Every model upgrade, every prompt edit, every new tool you connect can undo protection that was working last month.

This is also why the extra time is easy to waste. Sixteen more months on the high-risk rules sounds generous. Building a process that produces continuous evidence takes most of it. And when a regulator or an enterprise customer asks what you have, a year of test history is worth far more than a document written the week before.

Six things to do now

1. Update your dates. Anything planned around August 2026 needs rechecking.

2. Check your transparency basics. Are people told they are talking to AI? Is AI-generated content labelled? This is live today.

3. List your AI systems. What each one does, who it affects, and what protects it.

4. Work out which categories each one falls into (there may be more than one) and write down why. You will be asked.

5. If you generate images, video or audio, start on the December 2026 ban now. Test for the misuse, put safeguards in place, and keep the evidence. Alice's WonderBuild exists for exactly this kind of pre-launch testing.

6. Start producing evidence, not just plans. Run adversarial testing against your system now, keep the results, and record what your safeguards actually do once it is live. Alice's free risk check is a quick way to see where you currently stand.

How Alice helps

Alice handles the proving part: producing and keeping the evidence that shows your safeguards actually work.

Here is what that maps to, and when each one matters.

In practice, this is not three separate checks. It is one connected loop, and the loop is the point.

Red teaming finds how your AI breaks.

Automated attacks probe for prompt injection, model manipulation, data exfiltration and agents acting outside their intended scope. You get a full assessment: every attack run, whether it passed or failed, why, and how to fix it. That is the documented testing the law asks for, in a form your security and legal teams can sign off on.

Those findings become your guardrails.

They are not a generic filter bolted on top. Guardrails are formulated from your own red teaming results and trained on your exact policies, then run inline at low latency. Because they are built from real findings about your application, they block real risks instead of burying your team in false alarms.

Continuous testing keeps both sharp.

Red teaming keeps probing for new failure modes as models update and attacks evolve. Every finding updates your guardrails, and what those guardrails see in production sharpens the next round of tests. Both sides improve together, which is close to a literal description of what the law means by risk management that is ongoing rather than a one-time review.

Throughout, you get a record.

A trail showing what happened, when, and what your safeguards did about it. That is what you hand a regulator, an auditor, or an enterprise customer's security team.

Underneath all of it sits Rabbit Hole, our threat intelligence engine: more than a decade of knowledge about how malicious actors actually think and operate, now pointed at AI. It supplies the real-world attacks used to test your system, takes known attacks apart to generate new variations, and keeps updating as new tactics appear in the wild. It is the reason the tests reflect how AI genuinely fails rather than how a checklist imagines it might.

We spent a decade defending the platforms billions of people use every day against the most sophisticated bad actors online. That is why the leading foundation model labs trust us with their models, and why enterprises building on those models can now get the same protection.

What we do not do

We do not write your compliance documentation, run your conformity assessment, apply your CE marking, register your systems with the EU, or train your staff. Those are legal and process jobs, and you will need people or tools for them.

What we do is produce the evidence all of that paperwork is supposed to point at, and make your AI genuinely safer while we do it.

‍