Researchers have identified a linguistic unreadability vulnerability in LLM

A group of researchers published a work The Implications of Linguistic Illegibility for LLM Security, which proves the unreliability of self-reporting mechanisms of language models in security issues. The authors found that current protection systems, which rely on the interpretation of the model's text output, are vulnerable to bypass due to the fundamental inability of LLM to accurately assess its own internal states.

During the analysis, it was revealed that linguistic illegibility makes the model's attempts at self-control ineffective against targeted attacks. Researchers demonstrated that even with security instructions, the model cannot guarantee the absence of malicious code or hidden commands in its output, since its linguistic markers do not correlate with actual operations in memory.

To solve the problem, a transition from text filtering to taint tracking methods and strict process isolation is proposed. The work emphasizes that using external tools to verify output in 100% of cases is more reliable than trying to force the model to classify its own responses as "safe" or "dangerous".

The authors note that current architectures, which rely on prompt engineering to ensure security, have a critical failure threshold when encountering adversarial examples. The proposed isolation methods allow limiting the scope of code execution generated by the model, which reduces the risk of executing unauthorized instructions in the system environment.

This study changes the approach to designing AI systems, shifting the focus from attempts to "teach" the model security to implementing hardware and software restrictions at the execution level. This is critical for integrating LLM into infrastructure projects, where trust in text output is a point of failure for the entire system.