Go to PegaPoll
PEGAPOLL · NEWS

Google says attackers are using AI agents to find and exploit software flaws automatically

Cointelegraph · 2026-11-17
🤖 AI-generated content — The title and summary were produced automatically by artificial intelligence, without human editorial review.

Google Threat Intelligence Group (GTIG) says it has, for the first time, identified attackers using large language model frameworks to automate both the discovery and the exploitation of software vulnerabilities with no human directing each step. In a report published Monday, the researchers described a multi-agent vulnerability research and weaponization framework found on an exposed staging server. It was set up to run continuous differential analysis, comparing code before and after patch announcements on popular web applications, then to identify flaws and generate and execute exploit code on its own.

The approach turns N-day vulnerabilities, which are publicly known security flaws, into quick targets: by reverse engineering patches, attackers can weaponize a fix almost as soon as it is announced. Google warns that this shrinks the window organizations have to apply patches and lets attackers scale exploitation against internet-facing servers without human intervention. Its September report had stated that GTIG had not yet seen fully autonomous pipelines used against real targets.

Google also pointed to rising volumes. Vulnerability disclosures more than doubled between January and August. Exploited zero-days averaged eight per month in 2025 and 11 per month from January through August 2026, peaking at 22 in August. Automation also speeds up the move from initial access to domain administrator control; in an August 2026 PaperCut campaign, GreyNoise reported attackers reached that level within minutes.

The report linked two PHP vulnerabilities to LLM-assisted analysis. Cloudflare reported exploitation attempts against its customers in October, and technical details published before the Oct. 21 patch let researchers and attackers study the flaws. Anthropic separately warned in September of multi-agent operations automating most of the attack cycle, from discovery to data theft.

Continue in the app — vote & join in ➔
Source: Cointelegraph · via ahirlevel.hu

Related news