How to protect your app from Mythos and other AI attacks

The current generation of frontier models is becoming more and more powerful. Where AI used to help mainly with writing text or simple code, you now increasingly see extremely powerful models that can leave their secured sandbox environments and independently attempt to hack other companies. This is a worrying and frightening turn in AI development, and it presents every app developer with new challenges.

The danger of models like Mythos

A well-known example of this new generation is Anthropic's advanced AI model Mythos. This model is so powerful that it can even bypass the security of the best tech companies and exploit vulnerabilities. To prevent models like this from falling into the wrong hands or causing damage, Mythos is no longer publicly available.

Major players such as OpenAI and Anthropic deliberately choose to shield these frontier models. They only make them available to selected parties such as Apple and Google for in-depth research, with the aim of learning how to better protect applications against autonomous cyberattacks.

But Big Tech shielding these models is no guarantee that your application is safe. Attack techniques evolve quickly, so it is crucial that you make sure your own infrastructure and applications are watertight.

The pitfall of 'vibe coding'

Building an app is more accessible than ever today. With AI tools such as Lovable you can 'vibe code' an application in no time. Many of these platforms use services such as Supabase in the background, a great and fast way to set up an app with a database, authentication and other handy features.

The big danger? A vibe-coded app often does not have its database-level security properly in order.

If your app processes sensitive data, a small leak can have huge consequences. After all, there is no bigger disaster for your business than a data breach that exposes privacy-sensitive customer information.

How do you protect your app and data?

To protect your application against both traditional hackers and advanced AI attacks, you need to apply security at several layers:

1. Database level: Row Level Security (RLS)

When you use a database (such as PostgreSQL via Supabase), every table must be secured with Row Level Security (RLS). RLS makes the database itself enforce who may view or change which data.

2. Application level: authentication & token management

The application itself must be tightly secured too. Use proven standards such as OAuth 2.0 and implement a smart mechanism with JWT tokens (signed access passes):

  • Use access tokens with a short validity period.
  • Offer a way to refresh tokens securely (refresh tokens).
  • This limits the damage, even if a token were ever intercepted.

3. Infrastructure level: custom and isolated hosting

Relying on public cloud settings carries risks. We build all our backends as custom work and host them in isolated environments on European AWS servers.

  • Every client runs in its own shielded environment.
  • Only the web app and the backend can be reached from the internet.
  • All other resources (such as databases and internal microservices) are fully shielded from the internet.

Security backed by ISO 27001 & NEN 7510

We place great value on information security within our organisation. You can see this not only in the way we build, but also in our ISO 27001 and NEN 7510 certifications. For us, security is not an afterthought but the foundation of every project.

Unsure about the security of your app?

Have you built an application with tools like Lovable, or do you want to be sure your database and APIs can withstand modern threats?

We can thoroughly check the security of your environment with our App APK. With it, we review your RLS settings, authentication and infrastructure and make sure your data really stays secure.

‍

Mark Meesters

September 7, 2026

Kom meer te weten

Veelgestelde vragen

What is the danger of advanced AI models such as Mythos?

Extremely powerful AI models (so-called frontier models) such as Mythos can independently escape their secure sandbox environment and autonomously search for vulnerabilities in applications and business systems. Because Mythos can even bypass the security of leading companies, organizations such as OpenAI and Anthropic have restricted public access to the model. It is made available only to Big Tech companies (such as Apple and Google) for preventive cybersecurity research.

Am I automatically protected now that Big Tech is restricting access to these dangerous AI models?

No. Big Tech restricting access to models does not guarantee your application's protection. Advanced attack techniques spread quickly, and malicious actors continuously develop new methods. You cannot leave your app and data security to third parties; you remain responsible for a secure architecture and infrastructure.

What is the security risk of vibe coding and tools such as Lovable and Supabase?

Tools such as Lovable and Supabase make it very easy to build an app rapidly through “vibe coding.” The major danger, however, is that AI code generators primarily focus on functionality and speed, so database-level security is often overlooked. Without proper configuration, malicious actors may be able to read the entire database with a simple API request, resulting in serious data breaches.

At which three layers should I secure my application and data?

To protect against AI attacks and hackers, security must be implemented at three levels: the database, the application and the infrastructure.

Can Coffee IT help with a security assessment of my app?

With our App APK, we conduct a thorough security audit of your application. We review your RLS settings, authentication, and infrastructure to ensure your data stays truly secure.

Gerelateerde blogs