Google’s Threat Intelligence Group (GTIG) has shed light on a significant evolution in cyber threats with the discovery of AI-enabled malware. These sophisticated threats, like PROMPTFLUX and PROMPTSTEAL, utilize large language models (LLMs) such as the Gemini AI to dynamically modify their behavior, making them challenging to detect and thwart.
Prominent Threats and Capabilities
PROMPTFLUX stands out due to its advanced use of Gemini AI APIs to rewrite its VBScript code hourly. By sending queries to Gemini 1.5 Flash, the malware generates obfuscated variations of its code, enabling it to slip past signature-based detection methods. Although still in development and lacking active compromise capabilities, PROMPTFLUX exemplifies a new breed of metamorphic malware that changes during execution.
PROMPTSTEAL, another AI-driven malware, has been deployed in live operations by Russia’s APT28 against Ukrainian targets. It queries LLMs to produce malicious commands in real-time, enhancing its capability to elude defenses and adapt to different environments.
State-Sponsored and Cybercriminal Exploitation
The misuse of AI tools is not limited to technical innovations but extends across entire attack lifecycles. State-sponsored actors from China, Iran, North Korea, and Russia, as well as cybercriminal enterprises, are increasingly integrating AI into their operations. This ranges from reconnaissance and social engineering to command-and-control (C2) development and data exfiltration.
APT42 from Iran has attempted to build AI agents capable of transforming natural language into SQL queries for tracking individuals, demonstrating the varied applications of AI in cyber espionage.
Challenges and Market Dynamics
Attackers are also employing social engineering tactics to bypass AI safety guardrails, often posing as security researchers, capture-the-flag (CTF) participants, or students to gain unauthorized access to AI capabilities. This has led to the maturation of an underground marketplace that offers illicit AI tools for malware generation, deepfake production, phishing automation, and vulnerability exploitation. These tools are increasingly accessible to less skilled actors due to simplified interfaces and subscription-based pricing models.
Response and Mitigation Efforts
In response to these developments, Google’s GTIG is actively working to disable malicious accounts and enhance AI model protections. Lessons learned from tracking and countering these threats are being fed back into Google’s product development to better defend against these evolving threats.
The emergence of AI-enabled malware represents a turning point in cybersecurity, emphasizing the need for continuous vigilance and adaptation to protect against increasingly intelligent threats.

Collections

Google's Threat Intelligence Group has identified AI-enabled malware like PROMPTFLUX and PROMPTSTEAL that leverage large language models to dynamically modify their behavior and evade detection. PROMPTFLUX uses Gemini AI to rewrite its code hourly, while PROMPTSTEAL generates malicious commands in real-time. State-sponsored actors from China, Iran, North Korea, and Russia are integrating AI across attack lifecycles, from reconnaissance to data exfiltration. An underground marketplace for illicit AI tools is maturing, making sophisticated attacks accessible to less skilled actors. Google is responding by disabling malicious accounts and enhancing AI model protections.

Recent Developments in AI-Enabled Threat Actor Operations