Threat researchers report state-sponsored actors have developed AI-powered malware that can generate malicious scripts and alter its own code in real time to evade detection.

Examples include PromptFlux, a VBScript dropper that uses an AI partner to generate obfuscated variants, maintain persistence, and move laterally, effectively turning into an evolving metamorphic script.

AI-enabled malware families such as PROMPTFLUX, PROMPTSTEAL, and PROMPTLOCK dynamically generate or modify code and obfuscate behavior during execution.

State-sponsored groups, notably North Korea’s MASAN and PUKCHONG and Iran’s APT42, have been active in cryptocurrency theft, exploit development, and data-processing experiments that translate natural-language queries into SQL for personal data extraction.

Intersections of Iranian and China-associated actors include MUDDYCOAST, whose misstep exposed its command-and-control infrastructure, disrupting its operations.

Observed tools include QuietVault, a credential stealer using AI to hunt secrets; PromptSteal, a data miner that employs the Hugging Face API for exfiltration commands; and FruitShell, a reverse shell designed to evade LLM-powered defenses.

Gemini abuse spans the attack lifecycle across China-linked, Iranian, and North Korean actors, supporting phishing, vulnerability discovery, data processing, and C2; Google has disabled involved accounts and strengthened safeguards.

Threat actors pose as students or researchers in prompts to bypass AI safety guardrails and extract restricted information, including vulnerability or exploitation data.

These social-engineering techniques extend beyond productivity, enabling actors to manipulate AI to access restricted data.

Industry studies indicate more CIOs and COOs are adopting GenAI-driven cybersecurity solutions to reduce data security losses amid rising, sophisticated attacks.

The report advocates security standards for responsible AI use, with Google promoting the Secure AI Framework as a blueprint for secure lifecycle management.

PROMPTSTEAL, linked to APT28, queries open-source language models to generate Windows commands for file harvesting and data exfiltration.

State-sponsored actors from North Korea, Iran, and China misuse AI for reconnaissance, phishing lure creation, data exfiltration, and other malicious activities.