In its November 16, 2026 update, Google Threat Intelligence Group (GTIG) said it had identified a multi-agent framework configured to discover vulnerabilities and execute exploits without human intervention. GTIG called it the first fully automated workflow of this kind in its own observations. The framework, named exploit_agent, was hosted on an exposed staging server, according to GTIG and Mandiant’s report.
What exploit_agent automated
GTIG said exploit_agent used multiple frontier and local large language models (LLMs) to compare web-application software versions around patch announcements. It then analyzed potential vulnerabilities, drafted exploits and executed them automatically. The key change in this reported setup was linking those stages into one workflow without a person intervening between them.
GTIG also said the exposed server contained exfiltrated victim data and scripts for activity after a compromise. Its telemetry indicated interest in French websites in media, entertainment and education. Those observations describe what GTIG reported about this server and its telemetry.
Which other frameworks GTIG observed
GTIG reported seeing CAIRN and PentAGI deployed in the wild, while it saw experimentation with CyberStrikeAI and HexStrike AI. The report also described a customized PentAGI-derived platform used for reconnaissance and vulnerability scanning against organizations in Asia from late August to early October 2026.
GTIG says autonomous execution can increase speed and scale, but it can also leave forensic traffic, recognizable identifiers and traces in code. Those traces can create opportunities for defenders to detect activity. The report also cautions that purported LLM-generated proof-of-concept exploits can be nonfunctional, adding verification work for security teams.
Why system context matters for defensive triage
A vulnerability’s practical priority can change with the conditions around the affected system. In a Maze product walkthrough published December 4, 2025, security practitioner James Berthoty shows the platform assessing a vulnerability as non-exploitable in a test environment because its AWS network adapter lacked the required RDMA capability. In another example, Maze raises a medium-severity denial-of-service vulnerability in a Node.js server to high priority based on internet exposure and an available public exploit.
When you triage a vulnerability, check whether the affected system meets its prerequisites and is exposed; those details changed the assessment in Maze’s demonstration.
The walkthrough offers a separate defensive example: the relevant question for prioritization can include whether a system meets an exploit’s environmental prerequisites and whether it is exposed to the internet.