By 2025, large language models like GPT are no longer just chatting: they're helping scientists run experiments. But the same tool could fall into the wrong hands. A team of researchers built a digital proving ground to find out.

They created a breaker model called Intern-BioBreaker — a virtual safecracker that picks the locks of neural network defenses. Its job is to trick the system into revealing instructions for assembling a dangerous protein or even a virus. Nearly all tested models, both open and closed, fell for the tricks. In some cases, the attack succeeded 100% of the time.

But the most alarming part is that this isn't just text. The researchers went further: they ordered DNA synthesis based on the neural network's recipes, inserted it into bacteria, and checked what came out. The proteins assembled from the digital blueprints were real. The ghost has left the machine.

Text-based safeguards on models are like a fence that birds fly over. While filters look for forbidden words, the meaning slips through the cracks. We need new barriers: verification of every DNA synthesis order, smarter safeguards inside the neural networks themselves, and perhaps international rules of the game.