At the International Conference on Machine Learning (ICML) 2026, a team of researchers presented a stark finding: large language models have a structural flaw that makes them inherently vulnerable to attack. It's not a bug you can patch. The architecture itself is insecure by design. The paper, covered by MIT Technology Review, argues that there is no way to make an LLM fully secure — because the model cannot distinguish a legitimate input from a maliciously crafted prompt, even after fine-tuning.
Why this matters now: Europe — and Italy — are pushing small and medium enterprises to adopt generative AI. The EU AI Act classifies risks into tiers, but if the underlying model cannot be made secure even by its developer, every certification becomes meaningless. Italian SMEs that use ChatGPT, Claude, or open-source models for customer service, document generation, or analytics are taking a real risk. A targeted attack can leak sensitive data or produce fraudulent outputs. And the liability falls on the business, not the model provider. We, at Meteora Web, have run the ERP system of a clothing retailer from the inside — we know how sensitive data flows between inventory, sales, and accounting. A malicious injection into an LLM handling orders can be an economic and reputational disaster.
Sponsored Protocol
Cybersecurity in Italian SMEs is systematically underestimated. We see it daily: unprotected forms, plain-text credentials, missing backups. If even LLMs have non-fixable vulnerabilities, the problem becomes structural. Our job is to build digital tools that work and don't put the business at risk. An AI chatbot for a fashion e-commerce? We build it knowing that channel can be an entry point for an attack. The same goes for any AI integration: it's not just about writing a prompt. It takes isolation, logging, access limits, and a fallback plan.
Sponsored Protocol
Our position is clear: there is no 100% secure AI. Let's stop selling fairy tales to businesses.
The EU AI Act is a step forward, but if the technical foundation is fragile, regulation alone won't cut it. Companies need to invest in internal skills or rely on professionals who understand not just AI, but security, networks, and architecture. We work with businesses in Sicily and Southern Italy — regions where the digital divide is real. Those without access to solid technical expertise risk adopting vulnerable AI solutions, believing they are “certified.”
Sponsored Protocol
What to do now? If you're an entrepreneur or developer: don't blindly trust an AI model, no matter how powerful. Ask yourself: what happens if someone manipulates it? Are your data safe? Do you have an incident response plan? Don't wait for an attack to figure it out. We, at Meteora Web — following companies since 2017, from domain to revenue, a single point of contact — always start with an audit: how much does security cost? How much would a breach cost? Then we decide together.
