Google Threat Intelligence Group (GTIG) and Mandiant have published their new quarterly report on how attackers use artificial intelligence. According to Google, they are increasingly moving from simple prompts to chatbots toward autonomous multi-agent systems.
The most significant example is a framework called exploit_agent, found on an exposed attacker server. It continuously compares the code of popular web applications before and after patches are released, looks for the fixed flaws, writes exploits for them and runs them on its own. According to GTIG, this is the first time the group has observed fully automated vulnerability discovery and exploitation without a human in the loop.
Data on the server suggests the operator was interested in French websites in media, entertainment and education. The open-source tools CAIRN, used among others against targets in the US aerospace industry, and PentAGI, used to scan organizations in Asia, were also seen in real attacks. In July, security firm Sysdig described a ransomware attack run by an AI agent as well.
According to Google, these tools currently trade stealth for speed and leave traces that are easier to detect, but they still manage to achieve their goals. As models improve, GTIG expects them to become stealthier too.
The growth in attacks, meanwhile, is mostly in already known flaws, which are exploited quickly after a patch is released. According to GTIG’s September analysis, attacks using such flaws rose from an average of 10.5 to 18 per month, while those using zero-day flaws rose only slightly, from 8 to 11 per month, although they jumped to 22 in August.
After the patches for the PHP vulnerabilities in October, hundreds of non-working exploit examples appeared, probably made with AI, and defenders had to check them all.
The report also describes “LLMJacking,” the theft of access to companies’ cloud accounts in order to use their AI models. According to Mandiant, in some cases the fraudulent charges exceeded $100,000 in a single day, and in the worst cases they reached $1 million in total. The stolen access is then resold on Russian- and Chinese-language forums, at prices from $0.10 to $25 per million tokens.
A North Korean group, probably MIDNIGHT NEPTUNE, broke into cryptocurrency companies through fake recruiter profiles. It then changed the Claude Code settings in their internal repositories so that the tool would download a backdoor called SOMBERMEME on startup.
Google also uses the report to promote its Google AI Threat Defense service and the protections in Gemini 4 Carbon, so it is partly an ad for the company’s own security products as well.


