The Google Threat Intelligence Group on Monday said attackers are using autonomous artificial intelligence tools to find software flaws and build attacks without a person in the loop.

The finding is in the group’s latest AI Threat Tracker report, described in a post on Google’s blog. According to the report, bad actors now use autonomous large language model frameworks — software that can carry out tasks with a language model — to keep scanning software updates, work out which flaw a patch fixed, and build an attack for that flaw. The group said that sequence runs without human involvement.

Google said stopping threats that move at machine speed is not practical if defenders work only at human speed. In the post, the company wrote that “Defending against machine-speed threats at human speed isn’t feasible.” Google said it is offering Google AI Threat Defense for what it called the agentic era, when software agents act with less direct human control.

According to Google, the product combines its Gemini model and other models it called frontier models, context and priority rankings from Wiz, automated patching from CodeMender, and expertise from Mandiant. Google said those pieces are meant to keep finding and fixing flaws before attackers can use them. The post did not say what the product costs, who can buy it, or when it will be available.

The group directed readers to the full report on the Google Cloud Threat Intelligence blog. The post did not say how many attacks the report documented, which organizations were targeted, or which software was scanned.