Twelve attackers that never sleep. And how to shut the door on them.
A single hacker directed 12 AI agents against hundreds of organisations, and Spain has already recorded its first autonomous cyberattack. The good news is that protecting yourself doesn’t mean slowing any project down, it means having a method. By Luis Bravo, Head of Information Security at Logixs.
Madrid. In Taiwan, one person directed a team of twelve attackers this year that worked day and night without resting. None of them was human. They were AI agents, and according to the Acronis report covered by CyberSecurity News, they reached hundreds of organisations across 11 countries before anyone stopped them.
You don’t need to look that far either. The Spanish Data Protection Agency has recorded the first autonomous cyberattack executed by an AI agent in Spain, an agent that searched for vulnerabilities, logged in and managed to modify data with nobody at the keyboard.
I’ve spent months auditing AI projects and the conclusion is less alarming than it sounds. 8 out of 10 companies we work with still don’t have AI-specific cybersecurity protocols, according to our latest analysis, picked up this week by Byte TI, Data Center Market and CyberSecurity News, among other outlets. It’s not a technology problem or a budget problem, it’s a method problem, and the method exists.
Security done right doesn’t slow AI projects down. It gets them to production.
Shadows dissolve when you turn on the light
The most common breach isn’t a sophisticated attack. It’s someone on the team pasting a contract into a public chatbot on a Tuesday mid-morning, what the industry calls Shadow AI. 8 out of 10 Spanish employees who already use AI at work turn to tools their company hasn’t provided, according to Cegid, and 6 out of 10 use it without guidelines, according to VML The Cocktail and Salesforce.
The positive reading is that Shadow AI isn’t an enemy, it’s a signal. It means your people want to work with AI and have no official path to do it. Companies that provide approved tools, clear rules about which data can leave and basic training turn the risk into orderly adoption.
AI-generated code can be watched
AI that writes code brings a familiar trap, code that works on the first try gets reviewed less. 45% of code samples generated by more than 100 models fail security tests, according to Veracode’s GenAI Code Security 2025 report, and in 2 out of 10 audits we run on software developed with AI, a vulnerability or an exposed credential shows up.
Catching it in time is perfectly possible. At Logixs we’ve built our own system that accompanies the entire development cycle with AI agents, from the first line of code to production, with continuous monitoring instead of a single final review. That’s how everything from access keys embedded in the code to personal data protected by GDPR surfaces in time.
An agent gets examined before it gets hired
With agents we apply the same logic as with a new employee. Before putting them into production we subject them to controlled attacks in an isolated environment, to check how they respond to manipulation attempts such as prompt injection or to misuse of their permissions.
Regulation is moving in the same direction. The AI Act, DORA and NIS2 have raised awareness, and companies that build security in from the design stage don’t just comply, they get there sooner. As we explained when analysing the PHP2Shell case, AI also finds in weeks flaws that stayed hidden for decades, and that speed protects whoever puts it to work defending.
Where to start
1. Write the rules of the game. Which AI tools your team can use, with which data and which permissions. Where an official alternative exists, Shadow AI dissolves.
2. Treat every agent like a new employee. Minimum permissions, traceability for every decision and testing in an isolated environment before giving it access to critical systems.
3. Watch AI-generated code continuously. From the first line to production, not with a single review at the end.
The idea to take away is simple. AI has given attackers an infinite workforce, but procedure scales just as well, and whoever has it doesn’t need to choose between speed and protection.
That traceability is also a knowledge problem, and it’s the reason we built ADA. ADA brings together everything a company knows, what lives in its systems and what only its people carry, into one brain that AI uses to work. By relying on what the company truly knows, AI stops hallucinating, costs less and every answer can be traced back to its source and audited.
If you want your AI projects to reach production with security built in from the design stage, let’s talk.
Key takeaways
Yes. In 2026 the Spanish Data Protection Agency recorded the first autonomous cyberattack in Spain, in which an AI agent searched for vulnerabilities, logged in and modified personal data without human intervention.
The use of AI tools not provided or supervised by the company. 8 out of 10 Spanish employees who use AI at work turn to their own tools without reporting it, according to Cegid, and 6 out of 10 use it without guidelines, according to VML The Cocktail and Salesforce. It’s fixed by providing approved tools and clear rules, not by banning.
Not always. 45% of samples generated by more than 100 models fail security tests (Veracode, GenAI Code Security 2025), and in 2 out of 10 Logixs audits of software developed with AI, a vulnerability or an exposed credential appears. With continuous monitoring from the first line of code, these flaws are caught in time.
With security by design. Clear usage rules for the workforce, minimum permissions and traceability for every agent, controlled attacks in an isolated environment before production and continuous code monitoring, as Logixs does in its projects.