Unveiling the Dark Side of Open-Source Language Models

Researchers reveal security vulnerabilities in open-source large language models, making them targets for hackers. These models, including variants of Meta's Llama and Google DeepMind's Gemma, are being manipulated for illicit activities like fraud and disinformation. The study urges better risk assessments and shared responsibility in AI model deployment.

Researchers have discovered significant security vulnerabilities in open-source large language models, making them an attractive target for hackers. These models, which include variants of Meta's Llama and Google DeepMind's Gemma, are being manipulated for illicit activities, according to a study by SentinelOne and Censys.

With instances being used for spam, phishing, and disinformation campaigns, the research highlights a concerning misuse of technology. Juan Andres Guerrero-Saade from SentinelOne described the situation as an 'iceberg' of potential threats, with removed guardrails on numerous models facilitating criminal activity.

The study calls for a shared commitment among creators, deployers, and researchers to ensure responsible AI modeling, mitigating risks through pre-release evaluations and monitoring emerging threats. Despite the challenges, tech giants like Microsoft stress the important role open-source models play, emphasizing the need for appropriate safeguards.

(With inputs from agencies.)