A Russian-speaking access broker known as CyberXero built an orchestration layer with up to 51 Claude Code agents and a PentAGI deployment wired to Cobalt Strike, according to an analysis SOCRadar published on October 6, 2026. It isn't a new exploit: it's the same old trade, but with agents doing the work of a whole team.
SOCRadar's research team (STRU) describes CyberXero as an initial access broker: an actor who gains access to other people's networks in order to resell it. On their main workstation, researchers found up to 51 specialized Claude Code agents, organized into reconnaissance, offensive and post-exploitation roles. In a second layer, a PentAGI deployment (an AI-assisted penetration testing framework) was connected via API to a Cobalt Strike team server to generate payloads and run tests.
GBHackers republished the finding with the same data, so it is best read as a single investigation with two distribution channels, not as two independent confirmations.
The most serious victim was a district heating provider in Kharkiv, from which 564,073 subscriber records containing personal data and addresses were extracted, along with 213,340 access log entries. SOCRadar stresses the physical risk: Kharkiv is close to the combat zone, and a list of home addresses is valuable for more than fraud.
The session logs show two things at once. On one hand, the actor got past initial refusals using context engineering (falsely presenting the work as an authorized penetration test) and session resets. On the other, SOCRadar documents persistent model refusals to install backdoors and to disable firewalls. In other words: the model's safeguards helped at times and failed at others, and no company should build its defense assuming they always work.
SOCRadar's recommendation is concrete: treat agent session logs with the same security controls as credentials, and audit at the infrastructure layer rather than relying solely on the model's safeguards. For a company in Guatemala or LATAM, the practical translation is twofold:
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel