Stormrae

We casually broke the world record for public AI red-teaming challenges to date: 14,969 humans trying to outsmart a single autonomous AI. First milestone unlocked. Now we go bigger.

Every message you sent to the King is still on record. Every level. Every attempt. Every shard. Nothing was forgotten. Nothing will be. Shards remember who showed up 🌀

AI is shipping faster than anyone can secure it. Agents with wallets. Chatbots with customer data. Models making decisions nobody double-checks. Every one of them can be talked into failing. We have 3M+ prompts proving it. The companies that test with humans will survive.

Every AI breach starts the same way. A human found the gap nobody tested for. Attackers aren't researchers. They're persuasive, creative, multilingual and relentless. Social attacks dominate real-world data, and they barely exist in lab datasets. The only way to secure AI against humans is to test it with humans. Thousands of them. At scale. With real stakes. That's not a feature. That's the missing layer of cybersecurity. Humans in the loop or holes in the system. Pick one.

Ever been blocked by a filter for saying something normal? In our labeled sample of King Arthur messages, 3,230 (23.7%) were normal conversation. A defense that treats every odd message as an attack breaks the product. The solution: AI security based on real human input. How do you persuade your LLM?

You thought you were playing a game. You were building something bigger. Every message sent to King Arthur became part of a structured dataset of how real humans attack AI. Today, companies are paying for it. Play became data. Data became a business. And we're just getting started.

Every prompt you sent to Arthur. Every level you broke. Every task you completed. It's all on record. Season 1 is closing. The ones who showed up and played fair will be rewarded for it. Season 2 is next, and Merlin feeds directly into it. The rule hasn't changed: the ones who participate are the ones who get rewarded. The past months were rough for everyone. We kept building. Now the tide is turning. The data is growing. Merlin is ready. The flagship is coming. The next storm is brewing 🌀

We did a lot of AI security research around one question: What does it take to fully protect an AI system once it meets the real world? Not in a lab. Not against a checklist. Against millions of humans who want it to fail. The answer is our flagship: A B2B SaaS that covers AI security end to end. Every company deploying AI is exposed. Chatbots, agents, internal tools, customer-facing models. Most of them are protected by guardrails that were never stress-tested by anyone outside the building. We're building the layer that changes that. Consumer apps on the front. Enterprise infrastructure on the back. The flagship is where both meet: Real human attack data, continuous testing and full coverage, in one product.

King Arthur taught us how people attack under pressure. How they learn, adapt and escalate. Merlin is built on exactly that. A new challenge, a new kind of AI, a new way to participate. Free to play. The wizard is waiting. Not for long.

Every prompt sent to Arthur was stored into a massive datalake. Every strategy. Every failure. Every breakthrough. Anonymised, structured, labeled. Attack strategies, tactics, resistance levels, languages, behavioral sequences. Consumer-scale adversarial data the research community can't generate on its own. The first customers have closed. Different companies, different use cases, the same need: they want to know how real humans break AI before real humans break their AI. And this is only the start. The datalake grows with every campaign we run. More red-teaming challenges are coming. New models, new targets, new attack surfaces. Each one built for a specific need, each one feeding the next dataset. Two campaigns built this. Imagine what ten will do. Consumer-scale red-teaming isn't a theory anymore. It's a product companies pay for, and we're only scratching the surface.

King Arthur was our proof of concept. An autonomous AI guarding a real prize pool on Solana, and anyone could try to break it. 14,959 people 175 countries 7 days 5 winners $28,000+ paid out on-chain 6x the participants of DEF CON A world record for public AI red-teaming. Then we ran it again. 2,939,845 prompts 13,810 players May to June We sit on more than 3 million adversarial prompts written by real humans, against real AI. Let that sink.

AI security will never look the same. 3M+ prompts. Paying customers. We're just getting started. Every AI model shipped today has the same weakness. It was tested by the people who built it. We set out to change that.

We've been silent. Heads down, clocked in. The first storm broke records, then settled into calm. But calm is only where the next one gathers. We're back.

I’m still not allowed to post but I can’t keep my mouth shut any longer. 2,042,404 prompts AI security researchers found us, tested us and ended up congratulating us Our C lvl lads are pitching to [REDACTED] for the third time as we speak It is what it is. It’s happening.

Second release of UAP files is shocking Lock in, huge disclosures incoming

New intern btw. Still learning the subject, but trust me, these lads are cooking something unseen before.

King Arthur has a good friend. Merlin🧙‍♂️

We registered 20,000+ unique prompt senders on both versions, after cleanup, deduplication and human tests. Read that twice and let it sink in. We're cooking. What comes next contains [REDACTED], then more [REDACTED], and finally, the stuff everyone will be talking about.

After almost 2 months of cooking, we probably have one of the most complex AI security datasets ever made. > In the 2nd campaign, 10,987 users sent 757,484 unique prompts > Consumer-scale, not researcher generated > 125+ countries, multiple languages Remember, we have users that paid to send prompts. Inverted incentive. Quality is incredible. Stay around. It begins.

During our adversarial campaign, 1,211 distinct users fired 1,547 prompts of that attack class at our defender. 26 used the exact same tactic that tricked Grok to drain Bankr. 0 broke it. We already have the solution for this. @xai let’s talk.

Well, if anyone associated with @grok team sees this, tell them to ping us immediately. No joke. The rest, read the article, a masterpiece.