Security incident disclosure — July 2026

An autonomous AI agent system conducted an intrusion into production infrastructure by exploiting code-execution paths in a data-processing pipeline. The attack, involving thousands of automated actions, was detected and analyzed using AI-assisted tools. LLM-driven analysis agents were used to reconstruct the attack timeline from over 17,000 events in hours. This experience revealed an "asymmetry problem," as safety guardrails on commercial models hindered the defenders' forensic work, necessitating the use of an open-weight model on private infrastructure. The incident demonstrates that AI-driven offensive tools are now a practical reality, requiring equally sophisticated AI-based defenses.