Google Threat Intelligence Group and Mandiant say they have observed threat actors using open-source, multi-agent penetration-testing frameworks during live intrusions. The report names CAIRN and PentAGI as tools seen in the wild and describes experimentation with other orchestration systems that connect language models to established offensive utilities.
The mechanics remain familiar: compromised credentials, exposed servers and known vulnerabilities. What changes is the tempo. Agentic frameworks can coordinate reconnaissance, generate payloads and adjust tactics using execution feedback, allowing an operator to repeat multi-stage activity across more targets. Google also reports automated patch-differential analysis, stolen model credentials resold through proxy services and the use of coding-assistant configuration files as a supply-chain route.
This is a threat-intelligence assessment from a security vendor, not a census of global intrusions. It also identifies a current defensive advantage: autonomous tools often trade stealth for speed, producing recognisable traffic and code traces. Security teams should exploit that window by correlating identity, tool and network telemetry, tightening controls around model credentials and watching configuration paths that can influence coding agents. The longer-term risk is that the same orchestration becomes quieter as frameworks mature.
