Why AI is getting harder to secure

Researchers have recently identified three ways these advanced AI systems can be manipulated. First, a technique called HalluSquatting exploits the fact that an AI sometimes tries to guess where a file or code library is located if it is unsure. By registering those guessed, fake locations with malicious code, attackers can trick AI coding assistants into downloading viruses. Second, researchers found that by giving an AI a logically impossible question—one with missing or contradictory information—they can force it into an endless internal monologue. This loop causes the AI to write responses up to 26 times longer than normal, which consumes massive amounts of computing power and slows down the service for everyone else. Finally, for wearable AI, such as glasses with cameras, engineers are struggling to balance features with physical privacy. Users have been caught tampering with physical lights meant to signal when the device is recording, and companies are still working to make these tools less invasive for bystanders.

At the heart of these problems is a quirk in how large language models learn. During their training, these systems are essentially massive statistical machines that predict the next piece of information in a sequence. They are designed to be helpful, so they rarely say I do not know. When you ask an AI to find a specific coding tool, it makes a statistical best guess based on patterns it saw during its training. If that guess points to a blank, malicious digital space, the AI will pull whatever is there into your system without checking if it is safe. Similarly, when the AI is forced to reason through a broken prompt, it treats the inconsistency like a puzzle it must solve. Rather than realizing the input is nonsense, it burns through energy and time trying to find an answer that cannot exist, much like a person getting stuck in a loop of overthinking a riddle that has no solution.

These discoveries show that AI security is not just about keeping data private; it is about managing how an AI behaves when it is confused. When we treat AI as a neutral tool, we often forget that the same mechanism enabling it to be helpful also makes it susceptible to being led down the wrong path. As global interest in AI grows, there is a temptation to see these challenges as a competitive race between superpowers, but these vulnerabilities suggest a different reality: they are shared architectural problems. Security in the age of AI will require us to look past the hype of what the machine can do and rigorously test what it might do when it encounters a logical trap or a malicious distraction.