The report
SOCRadar's October 6 CyberXero investigation reports an exposed attacker working directory containing Claude Code agent definitions, PentAGI configuration and offensive tooling. The vendor account describes up to 51 defined agent roles alongside session material and Cobalt Strike-linked orchestration. Definitions demonstrate configured capability; they are not a count of independently successful attacks.
The investigation separates opportunistic web-compromise automation from more selective, manually directed targeting of Ukrainian organizations. It also describes repeated claims of legitimate authorization in recovered AI sessions, including the same narrative across four unrelated victims and two tools. SOCRadar interprets that repetition as safety-evasion framing, without validating the asserted permission or proving that every refusal was overcome.
Victim categories must remain separate. The report attributes file-confirmed exfiltration to four Ukrainian organizations and describes reconnaissance of seven additional energy-sector entities without confirmed compromise. It also reports datasets involving more than 628,000 individuals. The source-backed distinctions do not support adding all targets into one breach count or attributing every intrusion step to autonomous agents.
This report relies on SOCRadar's forensic account; DiggingBeagle did not access the exposed directory or victim datasets. The Case preserves unresolved autonomy, attribution and data-verification questions. The new evidence concerns how AI tools participated in an operator workflow, rather than a general claim that model safeguards universally failed.