Google Threat Intelligence Group (GTIG) said it had, for the first time, identified attackers using AI large language model frameworks to automate software vulnerability discovery and exploitation without step-by-step human direction. In the past quarter, the group said, it identified threat actors using such frameworks to fully automate these steps without a human in the loop, and it described this as the first time it had observed it. Google also said it has identified cybercriminals using AI agents to automate vulnerability discovery and exploitation without human intervention. Citing its latest AI Threat Tracker report, GTIG said bad actors are using autonomous LLM frameworks to continuously scan software updates, work out which flaw was patched, and build an attack for it, without human involvement.

The exploit_agent framework

GTIG identified an exposed staging server hosting a multi-agent vulnerability research and weaponization framework named exploit_agent. The framework performs continuous differential analysis of popular web application software, drafts exploits, and automatically executes them without human intervention. It compared software versions from before and after security patches were released and analysed code changes to identify potential vulnerabilities. It used multiple advanced, locally hosted AI models to scan popular web applications for security weaknesses. The server also contained scripts supporting post-compromise activities, including internal reconnaissance, password brute-forcing and webshell installation.

GTIG said the two PHP vulnerabilities involved in exploitation attempts were discovered using LLM analysis. Google also found data that had been exfiltrated from targeted victims.

Open-source tools and other groups

GTIG observed threat actors using open-source, multi-agent penetration-testing frameworks during live intrusions. The AI penetration-testing frameworks Google identified include CAIRN, PentAGI, CyberStrikeAI and HexStrike AI. GTIG has independently seen CAIRN and PentAGI deployed in the wild, and has seen experimentation with CyberStrikeAI and HexStrike AI. Between late August and early October 2026, investigators tracked a threat actor using a customised platform built with PentAGI for systematic reconnaissance and vulnerability scanning against organisations across Asia. GTIG also observed evidence that a PRC-nexus group was developing its own automated attack framework, likely for cyber espionage operations.

The PaperCut campaign

GreyNoise said a threat actor likely used LLMs to build and test exploits for the PaperCut vulnerabilities, assemble a target list, and run a widespread campaign against hundreds of organizations four days after initial disclosure. The attacker reportedly expanded an initial foothold to domain administrator within minutes. The GTIG report cited an August 2026 PaperCut campaign in which attackers reached domain administrator access within minutes, according to GreyNoise. Researchers and the cybersecurity firm Huntress generated working exploits for the PaperCut vulnerabilities within a day of disclosure. AI involvement in the GreyNoise-described campaign remains an assessment rather than a definitively established fact.

Limits and defenses

Google says it has not yet seen a fully autonomous attack pipeline turned loose on real targets. In its September report, GTIG said it had not yet observed threat actors deploying fully autonomous pipelines against targets in the wild. Google warned that AI-generated exploits can fail when tested, and that automated frameworks may leave traces that security teams can detect. It also warned that automated AI workflows give organizations less time to apply security patches before attackers develop ways to exploit vulnerable systems.

According to Google, zero-day vulnerabilities exploited averaged eight per month in 2025 and 11 per month from January through August 2026, reaching 22 in August 2026. Google is offering a product called Google AI Threat Defense, which combines Gemini and other frontier models, Wiz's context and prioritization, CodeMender's automated patching, and Mandiant's expertise. Google says the product is meant to continuously find and fix flaws before attackers can exploit them.