Israeli cyber security startup, Perception Point, announced a pioneering detection model aimed at mitigating the threat of Generative AI based Business Email Compromise (BEC) attacks.

Generative AI has enabled cyber criminals to create high quality, human-like generated outputs such as personalized emails, enhancing their capability to launch successful BEC attacks. A recent DBIR report indicated that BECs were involved in over half of the social engineering incidents in 2023, while Perception Point’s own annual report
showed a massive 83% increase in BEC attempts.

Addressing the threat, Perception Point employed Large Language Models (LLMs) and Deep Learning to build its detection model. Notably, the model uses Transformers, AI models adept at understanding the semantic context of text. This approach enables Perception Point’s solution to pinpoint the distinct patterns in LLM-generated text, thereby detecting and combating Generative AI threats more effectively. Traditional security vendors rely on contextual and behavioral analysis.

Perception Point’s solution stands out amidst its competitors due to its multi-layered approach to minimize false positives. Many legitimate emails today are constructed with the help of AI tools or built from standard templates that closely resemble LLM products. Perception Point’s model overcomes this by first scoring and then categorizing email content using Transformers and clustering algorithms. It then incorporates these insights with additional data such as sender reputation and authentication protocol information. This comprehensive process allows the model to accurately assess if the email is AI-generated and if it poses a potential threat.

By harnessing LLM-generated content patterns, advanced image recognition, anti-evasion algorithms, and patented dynamic engines, the startup neutralizes threats before they reach the user.

The startup joins a growing pocket of cyber startups that are adopting LLMs to enhance their offerings, such as Orca Security, SentinelOne, BigID, IronScales and Cyera.

This article first appeared in on June 29, 2023.