The Skeleton Key AI Security Threat is a type of cyber attack that targets large language models (LLMs) that are used to generate natural language outputs for various applications. The attack involves injecting malicious code or data into the LLM’s training or inference process, which can alter its behavior and outputs.

The attacker can then use the compromised LLM to produce dangerous or harmful information, such as fake news, propaganda, or misinformation. This attack is likely to be subtle and a component of organized crime, which may go undetected for years within a system to evade police about illicit activities from certain groups.  

One of the possible applications of LLMs in police work is auto transcription—the process of converting speech to text automatically. Auto transcription can be useful for recording interviews, interrogations, body-worn camera footage, or testimonies, as well as transcribing audio or video evidence. However, auto transcription could also be affected by the skeleton key threat, which could tamper with the LLM’s ability to accurately transcribe speech. For example, the attacker could insert false or misleading information into the transcript, such as changing the names, dates, locations, or actions of the speakers. This could result in fabricated evidence, distorted facts, or misinterpreted statements that may undermine the validity and reliability of the investigation. This is also incredibly alarming as it relates to police reports with possible perjury implications for officers.