State-Backed Hackers Leverage Gemini AI for Recon and Attack, Google Reveals

Taylor Keatsman

Published: February 13, 2026 8:32 AM ET

State-Backed Hackers Leverage Gemini AI for Recon and Attack, Google Reveals

Recent findings from Google reveal that state-backed hackers, specifically a group linked to North Korea named UNC2970, are employing the generative AI model Gemini for reconnaissance and cyber attacks. This threat actor is utilizing Gemini to conduct advanced profiling of high-value targets, facilitating more effective cyber campaigns.

Gemini AI Used for Cyber Operations

The Google Threat Intelligence Group (GTIG) reported that UNC2970 has employed Gemini to synthesize open-source intelligence (OSINT). Their activities have targeted major cybersecurity and defense firms, mapping critical job roles and associated salary information.

Techniques and Targeting

This strategy represents a shift, blending professional research with malicious reconnaissance. By crafting tailored phishing personas, UNC2970 has become adept at identifying vulnerable targets for cyber intrusions.

Group Name: UNC2970 (North Korean hacking group)
Related Groups: Lazarus Group, Diamond Sleet, Hidden Cobra
Key Focus: Aerospace, defense, energy sectors
Operation: Dream Job

Adoption of Gemini by Other Threat Actors

UNC2970 is not alone in utilizing Gemini. Other hacking groups leveraging this AI include:

UNC6418: Targeting sensitive accounts.
Temp.HEX (Mustang Panda): Gathering intel on individuals in Pakistan.
APT31 (Judgement Panda): Automating vulnerability analyses.
APT41: Extracting code explanations from open-source tools.
UNC795: Researching and developing web shells.
APT42: Crafting engaging personas for social engineering.

Concerns Over Generative AI Misuse

Threat actors are increasingly reframing their prompts to exploit AI models like Gemini, pretending to be security researchers. Steve Miller, AI threat lead at GTIG, emphasized the ongoing efforts to enhance safety measures against such misuse.

Emerging Malware Solutions

Google has also highlighted the detection of several malware variants, notably HONESTCUE, which utilizes the Gemini API. This malware acts as a downloader and can execute code in memory without leaving traces.

Phishing Kits and ClickFix Campaigns

Additionally, a phishing kit dubbed COINBAIT has surfaced, masquerading as a cryptocurrency exchange. This kit has been linked to the threat group UNC5356. Recent ClickFix campaigns have also been identified, which lure users into downloading information-stealing malware.

Model Extraction Attacks

Google reported active model extraction attacks aimed at replicating its proprietary AI models. In a notable incident, over 100,000 prompts queried Gemini to extract information, underscoring vulnerabilities in AI privacy. According to researcher Farida Shafik, protecting model weights alone is insufficient. The behavior of a model is exposed through its interactions with users.

Proactive Measures in AI Cyber Defense

In response to these challenges, Google has established the AI Cyber Defense Initiative. This program, launched in 2024, aims to leverage AI to enhance cybersecurity defenses, acknowledging that adversaries are equally capitalizing on AI technologies.

As AI continues to evolve in both offensive and defensive capacities, organizations must invest in AI-driven defensive frameworks to stay ahead of emerging threats.