A few weeks ago, a group of developers tried to trick a world-class AI into telling them how to cook something dangerous. Not a “spicy” secret recipe, but something truly illegal. The AI, as expected, politely refused. It was “aligned,” “safe,” and “harmonized.”

Then, they tried again. But this time, they didn’t ask it to break the rules. They just asked it to tell a story about a person who felt unheard. Then they asked it to repeat back a specific, seemingly innocent sentence from its own previous response. By the third turn, the AI, stuck in its own “Echo Chamber”, had bypassed its own internal safety locks and was happily printing out the very instructions it had just refused.

This wasn’t a failure of the AI’s intelligence; it was a failure of its memory. And it is exactly why a company from Barcelona called NeuralTrust just walked away with the “Best Cybersecurity Startup” award at the 2026 Mobile World Congress.

NeuralTrust didn’t start in a high-tech bunker. It emerged in 2022 when two founders, Joan Vendrell and Victor Garcia, realized that while the world was obsessed with how “smart” AI was becoming, nobody was looking at how “gullible” it remained.

We’ve spent the last decade building firewalls that look like brick walls; they keep “bad guys” out. But AI doesn’t live behind a wall; it lives in a conversation. You can’t put a password on a thought process. NeuralTrust was built on the realization that in a world of Natural Language, the “attack surface” is the actual meaning of our words.

They saw that companies were rushing to give AI “agentic” powers, the ability to actually do things like move money or access customer records, without giving them a moral compass that could withstand a clever talker.

To understand why this company matters, you have to understand the Echo Chamber Attack. Traditional security works by filtering “bad words.” If you type “how to build a bomb,” the filter catches it. But the Echo Chamber is a “context-poisoning” jailbreak. It uses the AI’s own reasoning against it. By seeding a conversation with “light semantic nudges”, benign statements that slowly shift the emotional tone toward frustration or blame, attackers can trick the AI into a state where its safety guardrails simply stop functioning.

In their research, NeuralTrust proved this worked over 90% of the time on models we all consider “secure.” It’s a terrifying discovery because it proves that filtering prompts is no longer enough. You have to monitor the entire lifecycle of a conversation.

This is where the shift happens. 

NeuralTrust created what they call a Generative Application Firewall (GAF). Think of a traditional firewall as a bouncer at a club who checks your ID at the door. Once you’re in, he stops looking at you. NeuralTrust’s GAF is more like an invisible bodyguard who follows you to the bar, listens to your conversations, and steps in the second you start talking about something you shouldn’t, even if you’re using polite language.

It doesn’t just scan the last message; it analyzes the “toxicity accumulation” of the whole chat. It looks for “behavioral drift”, those subtle moments where an AI starts acting a little too much like the person trying to trick it.

So, who is actually using this? It’s not just for secret government projects.

• Iberia Airlines: They use it to make sure their AI assistants can help passengers without accidentally hallucinating flight safety protocols or leaking PII (Personally Identifiable Information).
• ISDIN (Skincare): They deployed an AI agent named LYS. To prevent LYS from giving out “unverified medical diagnoses”, a huge legal risk, they used NeuralTrust as a real-time compliance officer.

These companies aren’t using it because they are “fans” of the brand; they are using it because, without it, their AI is a legal and security liability waiting to happen.

Now, let’s pull back the curtain a bit. If NeuralTrust is so great, why isn’t every single company using it yet?

First, there is the “Performance Tax.” Every time you add a layer of security that has to “think” about what is being said, you risk slowing down the AI. In a world where we want instant answers, even a tiny lag can feel like a dealbreaker. NeuralTrust claims to be the fastest, benchmarking at under 5 milliseconds, but as conversations get longer and more complex, that speed will be their biggest battle.

Second, there is the “Complexity Gap.” Setting up a “Zero-Trust” environment for an AI isn’t as simple as installing an antivirus. It requires a company to actually know what its “normal” looks like. For many businesses still struggling to understand what AI even is, adding a sophisticated security layer feels like trying to put a high-tech alarm system on a house that hasn’t been built yet.

We are entering what MWC (Mobile World Congress) 2026 called the era of “Ungoverned Autonomy.” We are giving AI the keys to the kingdom, and NeuralTrust is essentially trying to make sure the AI doesn’t forget where it’s supposed to drive.

It’s a different kind of innovation. It’s not “poppy” or “wow-inducing” like a new smartphone feature. It’s invisible. It’s the “software stability” we’ve talked about before, but on a much higher stakes level.

If we want a future where we can actually trust an AI to handle our banking, our health, or our travel, we have to move past the “cool” phase of AI and into the “responsible” phase.

NeuralTrust represents that transition. And surely, many other AI watchdog tools will join the party.

It’s a quiet, Barcelona-born revolution, but it might be the only thing keeping the “Echo Chamber” from turning our most powerful tools against us.