CyberStrikeAI appears in observed education-sector credential theft and database export

The investigation separates an operator-controlled command-and-control system from its endpoint implant. In one reconstructed timeline, access to a SQL Server host preceded the addition of the CyberStrikeAI implant by roughly two weeks. Later activity included directory reconnaissance, remote execution and credential collection before database records were exported.

The evidence consists of incident telemetry and recovered framework behavior, not a full synchronized record of every model prompt and human command. Chinese-language artifacts and local time-zone indicators cannot establish an operator nationality, and the number of affected organizations is not fully reported.

The recovered architecture separates endpoint capability from model decision-making. A Go implant on a compromised host handled communications and remote tool execution; the agentic planner was hosted in the operator's controller, which exposed commands through MCP. Kaspersky's telemetry documents reconnaissance, credential access, lateral movement and database export in real education-sector intrusions, but it does not provide synchronized prompts proving which steps the model chose without human intervention. The separate CyberStrikeAI campaign reported in Japan should retain its own victim and actor identity.