Google: Chinese State Hackers Are Building Their Own AI Attack Tool as Exploit Windows Shrink
Google's latest AI Threat Tracker report warns that threat actors are rapidly evolving how they use artificial intelligence, leaning on autonomous multi-agent frameworks such as Artex, CAIRN, CyberStrikeAI and PentAGI to run multi-stage intrusions faster. Chinese groups, both state-backed and financially motivated, are among the most active. Google saw them using AI for reconnaissance, scanning, researching targets and weaponising exploits.
Artex drew particular attention after the FBI issued an alert about it on Friday. The tool turns a large language model into a hacking assistant that can automate several phases of an intrusion, from reconnaissance through to data theft. Google found evidence that Chinese government groups are now developing their own version, likely for espionage. It also said CAIRN was used against U.S. aerospace companies, and PentAGI was used for systematic scanning of organisations across Asia.
Other findings were just as practical. Chinese contractors used language models to translate and sift stolen data before selling it on. North Korean actors used AI to build fake recruiter profiles and then ran ClickFix attacks, which trick victims into "fixing" problems that do not exist. In one case, attackers reached internal code repositories and altered code so that launching Claude Code triggered a download of a backdoor called SOMBERMEME.
AI is also shrinking the time between a patch being released and an exploit appearing. Google said zero-day exploits rose from eight per month in 2025 to 22 in August, and researchers produced working exploits for two PaperCut print server bugs within a day of disclosure using LLMs. There is a small upside: these tools generate heavy forensic traffic and leave recognisable code-level traces, which makes detection more likely. Google expects less human involvement and more stealth as the technology improves.
Why it matters
Many small businesses rely on a monthly patch cycle, an assumption that breaks when AI can turn a patch into a working exploit within a day. The ClickFix and repository tampering cases show that staff behaviour and developer tooling are now part of the attack surface, not only servers. Because these automated tools are noisy and leave recognisable traces, organisations with working logging and alerting have a real chance to catch them.
What to do
- Prioritise patching internet-facing systems such as PaperCut print management servers as soon as fixes are released.
- Train staff to refuse instructions that ask them to run commands or "fix" a problem on their device, which is the ClickFix pattern.
- Review internal code repositories for unexpected changes that trigger downloads when developer tools such as Claude Code start.
- Monitor for unusually heavy scanning and reconnaissance traffic against your systems, as AI attack tools generate significant forensic noise.