Google Reports Hackers Attempt Gemini Model Extraction and AI-Powered Attacks
Google's Threat Intelligence Group reported that attackers are attempting to extract and replicate Gemini's capabilities, while government-backed groups and malware developers increasingly integrate AI into cyberattacks for social engineering, reconnaissance, and to evade detection systems. The threats include model extraction attempts by private firms and academics, state-sponsored actors using AI for targeted attacks on defense sectors, and malware families enhanced with AI functionality.
What happened
Google Threat Intelligence Group (GTIG) and Google DeepMind detected actors attempting to obtain information about how generative AI models reason in order to replicate their functioning. The majority of these extraction attempts came from private companies and academic researchers, with Gemini's reasoning capability being a frequent target.
GTIG identified instances of malware integrating AI capabilities, including HONESTCUE, which used Gemini's API to generate functions while attempting to evade traditional detection systems. A phishing kit called COINBAIT was also created with the aid of AI to mimic a major cryptocurrency exchange and steal credentials.
Reported impact
Government-backed groups including APT42, linked to Iran, used AI to locate official emails and gather information to create credible social engineering attacks. UNC2970, linked to North Korea, used Gemini to analyze public information and profile high-value targets in defense sectors by impersonating company recruiters.
Index analysis
The report indicates that AI is being used throughout all phases of attacks, particularly for reconnaissance and social engineering. GTIG observed that attackers are increasingly integrating AI into malware families and accelerating fraud industrialization through AI-powered code generation tools.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.