technology

Google Warns: CCP Hackers Using AI for Automated Cyber Attacks

Published November 20, 2026 · 5 min read
An engineer from the Israeli company “Commun.it” uses his expertise in social media commercial analysis to identify networks of fake users during at the group’s office in the Israeli city of Bnei Brak near Tel Aviv on January 23, 2019. – A coalition of Israeli diplomats, programmers and hackers have joined forces to stave off threats — including from hostile states — by identifying networks on social media and getting them removed. (Photo by JACK GUEZ / AFP) (Photo credit should read JACK GUEZ/AFP via Getty Images)

A recent report from Google, the latest edition of the “AI Threat Tracking Report,” highlights concerning trends regarding the use of artificial intelligence (AI) tools by state-sponsored hacker organizations and profit-driven cybercriminal gangs within China. These groups are reportedly engaging in a wide array of cyber activities, including espionage, vulnerability scanning, data theft, and manipulation of public opinion.

In addition to Chinese hackers, the report indicates that cybercriminals from Iran and North Korea are also employing AI in their cyber offensives. These activities range from creating fake job profiles to deploying malware and attacking cryptocurrency firms. Google warns that the incorporation of AI into these operations is likely to make future cyber attacks faster, stealthier, and more complex.

According to media sources citing the report, while Iran and North Korea are utilizing AI in their cyberattacks, the scale and sophistication of their efforts pale in comparison to those of the Chinese state-sponsored hackers. Google’s researchers have uncovered a multitude of malicious activities that these hackers are orchestrating using AI tools, including reconnaissance, front-end scanning, and the weaponization of software vulnerabilities targeting various sectors and regions.

The findings reveal that Chinese cyber contractors are employing large language models (LLMs) to translate and organize vast amounts of stolen data, which is then categorized and sold to clients. Some hackers in China are also leveraging AI-hosted document storage infrastructure to hold numerous phishing bait files, using these platforms to transfer stolen data.

Moreover, multiple Chinese hacking organizations are reportedly testing the Artex tool on a large scale. The FBI issued a warning about this development last Friday, raising alarms among cybersecurity experts. Developed by a Chinese entity, Artex allows users to transform large language models into powerful hacking tools. The FBI points out that Artex can automate multiple phases of cyber intrusion, from initial reconnaissance to data theft and transfer, at machine speed.

Additionally, Google has discovered that Chinese hacker groups are developing their own versions of tools similar to Artex for espionage activities. Reports indicate that these hackers are either using existing tools like CAIRN, PentAGI, CyberStrikeAI, and HexStrike AI in real attacks or testing them privately.

The CAIRN tool enables multiple AI agents to collaborate in launching cyber intrusions against specific targets. This has been evidenced by hackers utilizing CAIRN to penetrate aerospace companies in the United States. Meanwhile, other hackers have utilized PentAGI to conduct systematic reconnaissance and vulnerability scanning against organizations across various Asian industries.

Google’s research indicates a widespread use of large language models by adversarial states such as China, Iran, and North Korea, along with cybersecurity providers. This usage encompasses scanning for software vulnerabilities, generating false content, manipulating search engine optimization (SEO), and producing material aimed at shaping public opinion.

For instance, the Iranian hacking group APT42 has used multiple large language models, including Google’s Gemini, to carry out targeted reconnaissance and social engineering attacks. Similarly, North Korean hackers have crafted job resumes with AI assistance, which they then employed during interviews to help answer professional questions. In cases involving recent cyber breaches in the cryptocurrency sector, North Korean hackers created fake job resumes and then utilized those false identities to perpetrate ClickFix attacks.

This method of attack involves hackers deceiving victims into believing their devices have malfunctioned, prompting them to perform specific actions to resolve the fabricated issues, thereby gaining unauthorized access to the devices.

North Korean hackers have employed this tactic to attack several cryptocurrency firms. This method was also observed in a long-term cyber attack campaign recently dismantled by the FBI.

Google has identified another variant of this attack technique where North Korean hackers use ClickFix to infiltrate victims’ devices, subsequently progressing to breach the internal code repositories of targeted enterprises. They implant malicious code directly into the large language models used by these organizations. When users activate the software known as Claude Code, a backdoor program called SOMBERMEME is automatically downloaded.

Furthermore, hackers are utilizing large language models to automate the previously labor-intensive process of software vulnerability analysis. This enables them to rapidly analyze software patches and identify vulnerabilities, significantly shortening the window of time between a vulnerability’s disclosure and exploitation by hackers.

According to Google, the number of zero-day exploit attacks has seen a remarkable increase this year, rising from an average of eight instances per month in 2025 to 22 by August of this year.

The report includes examples such as two software vulnerabilities identified in the PaperCut print management server over the summer. Researchers employed large language models that enabled them to develop operational attack programs within a day of the vulnerabilities being disclosed.

Researchers highlighted that while AI tools significantly reduce the time required for cyberattacks, allowing some hackers to execute attacks at “machine speed,” they also generate a substantial amount of digital activity logs, which can facilitate forensic analysis. Moreover, these tools tend to rely on easily recognizable features, leaving clear traces at the code level.

While Google notes that these traces could enhance the likelihood of detecting and thwarting cyberattacks, it also emphasizes the considerable potency of these tools in improving attackers’ efficiency.

Lastly, Google cautions that as these tools and their underlying large language models evolve in complexity, the role of human involvement in future cyber-attack operations is likely to diminish, making attacks even more covert and sophisticated.