Threat actors are increasingly abusing artificial intelligence (AI) tools for cyber operations, moving beyond simple phishing lures and malware coding to more advanced activities such as exploit development and autonomous attack orchestration. According to a recent report from Google's Threat Intelligence Group (GTIG), adversaries are leveraging large language models (LLMs) to discover vulnerabilities, develop exploits, and automate multi-stage attacks with minimal human oversight.
Since the widespread availability of LLM-based tools, cybercriminals and state-sponsored actors have integrated AI into their workflows. While earlier abuses included generating more convincing phishing emails or writing malicious code snippets, the latest findings indicate a significant escalation. GTIG identified a threat actor that likely used an AI model to develop a zero-day exploit—an industry first, according to the researchers.
AI-Assisted Exploit Development
The exploit in question targets a popular open-source web-based system administration tool, allowing an attacker to bypass two-factor authentication (2FA) using valid user credentials. The vulnerability was implemented in a Python script that contained telltale signs of AI generation: an abundance of educational docstrings, a hallucinated CVSS score, and a clean, structured format characteristic of LLM training data. Although the actor planned to use the exploit at scale, GTIG disclosed the bug to the vendor to disrupt potential attacks.
This development is part of a broader trend where adversaries use LLMs for vulnerability research. Google observed suspected Chinese actor UNC2814 prompting Gemini to act as a network security researcher auditing embedded devices, such as TP-Link firmware, for pre-authentication remote code execution vulnerabilities. Similarly, the North Korean group Silent Chollima (APT45) sent thousands of repetitive prompts to analyze different CVEs and validate proof-of-concept exploits, thereby strengthening their exploit capabilities. Some actors have also trained models on specialized vulnerability repositories like "wooyun-legacy," which contains over 85,000 real-world vulnerability cases collected between 2010 and 2016. Additionally, threat actors are experimenting with agentic tools like OpenClaw and OneClaw to assist in vulnerability discovery.
AI-Powered Attack Orchestration
One of the most notable use cases detailed in the report involves the Android backdoor "PromptSpy," first tracked by ESET. This malware abuses Gemini to keep itself in the recent apps list and uses AI to navigate the Android user interface autonomously. It interprets real-time user activity and can capture biometric data to replay authentication gestures, regaining access to a compromised device. This represents a shift from traditional malware to AI-driven orchestration that adapts to user behavior.
Beyond individual malware, threat actors are deploying agentic workflows to execute complex, multi-stage attacks. A China-nexus actor used agentic tools such as Hextrike and Strix in attacks against a Japanese technology firm and an East Asian cybersecurity platform. These tools maintain persistence across the attack surface, automate vulnerability verification, and reduce the need for human intervention. GTIG noted that this combination of autonomous reconnaissance and automated verification marks a transition toward AI-driven frameworks capable of scaling discovery activities with minimal oversight.
This trend mirrors the evolution in cybersecurity defense, where organizations are moving from human-in-the-loop to human-on-the-loop models—AI agents make moment-to-moment decisions while humans intervene only when necessary. John Hultquist, chief analyst at GTIG, warned that defenders must incorporate AI into their defenses or risk facing "machine time threats at human speed." He emphasized that without AI, security teams will be overwhelmed by a deluge of alerts and incidents, unable to keep up with adversaries that operate faster than patch cycles and move laterally across networks.
The use of AI by threat actors is not limited to state-sponsored groups; cybercriminal gangs are also adopting these technologies. For instance, Latin American threat actors have been observed generating custom hacking tools on the fly using AI. The accessibility of LLMs has lowered the barrier to entry for sophisticated attacks, enabling even less skilled actors to develop and deploy advanced exploits.
As AI models like Anthropic's Claude Mythos demonstrate the ability to find critical zero-day vulnerabilities through natural language instructions, the security ecosystem must prepare for a new era of AI-driven threats. While GTIG's report does not suggest that current adversaries are using such advanced models, the trends indicate that AI will increasingly shape both offensive and defensive cybersecurity strategies.
Source: Dark Reading News