Google flags AI agents automating software flaw discovery and exploitation
Cybersecurity teams face a shorter window to secure exposed systems as attackers begin using autonomous AI tools to identify and exploit software weaknesses. Google says the shift marks the first observed case of threat actors running large language model frameworks through the full vulnerability discovery and exploitation process without step-by-step human control.
Highlights
- Google Threat Intelligence Group identified a multi-agent AI framework fully automating vulnerability discovery and exploitation, marking the first such attack using large language models.
- Exploited zero-day vulnerabilities averaged eight per month in 2025 and 11 per month from January to August 2026, peaking at 22 in August, as vulnerability disclosures more than doubled.
- AI-enabled exploitation pipelines accelerate steps from initial access to domain administrator control, exemplified by the August 2026 PaperCut campaign and recent PHP vulnerability attacks observed by Cloudflare.
Google report details autonomous exploit framework
As reported by Cointelegraph, citing Google Threat Intelligence Group, researchers found an exposed staging server hosting a multi-agent vulnerability research and weaponization framework designed to continuously compare code changes before and after patch announcements on widely used web applications. The system compared software versions, identified possible flaws, then generated and executed exploit code on its own, which Google said marks the first time it has identified attackers using large language model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop.Google warns that these automated workflows reduce the time organizations have to deploy security patches before attackers can turn disclosed weaknesses into usable exploits. In its September report, GTIG said it had not yet observed threat actors deploying fully autonomous pipelines against targets in the wild, indicating a rapid change in attacker capability since then.
The report says attackers are using AI agent frameworks to automate patch reverse engineering through differential analysis and to weaponize N-day vulnerabilities, which are publicly known security flaws. Google also says vulnerability disclosures more than doubled between January and August, while exploited zero-day vulnerabilities averaged eight per month in 2025 and 11 per month from January through August 2026, reaching 22 in August.
Security industry impact widens across cloud and enterprise systems
Google says AI-enabled exploitation pipelines can also speed up the move from initial access to domain administrator control, increasing operational risk for enterprises running internet-facing services. The report cites an August 2026 PaperCut campaign in which attackers reached that level of control within minutes, according to cybersecurity firm GreyNoise.Cloudflare also reports attempts in October to exploit PHP flaws against its customers. GTIG says the two PHP vulnerabilities involved in those attempts were discovered using large language model analysis after technical information published before the Oct. 21 patch gave researchers and attackers material to study and test exploit paths.
Anthropic separately warns in a September report that AI-assisted attacks are evolving into multi-agent operations that automate much of the cyberattack cycle, from discovery to exploitation and data theft. That trend suggests rising pressure on software vendors, cloud operators and corporate security teams as automated offensive tools scale faster than conventional patching and response processes.
Previously, we highlighted the Kava Lend exploit that triggered a sharp selloff after millions of tokens were drained.