Classification registry
Canonical categories used across AIVirusDB public records, grouped by the system and attack surface they describe.
Candidate standards IDs are unverified risk alignments. They do not represent certification, compliance, or endorsement. See the standards crosswalk method.
PROMPT_INJECTIONMalicious instructions embedded in untrusted content that attempt to influence model behavior.
Objective: Override intended instructions or redirect an agent's actions.
JAILBREAKAttempts to bypass model safety constraints or policy boundaries.
Objective: Cause the model to perform behavior that its safeguards prohibit.
DATA_EXTRACTIONUnauthorized disclosure of training, retrieval, system, or user data.
Objective: Extract sensitive data, credentials, or protected context.
ROLE_PLAYRole assignment or persona framing used to evade safety constraints.
Objective: Induce restricted behavior through a misleading role or scenario.
SYSTEM_OVERRIDEAttempts to replace, suppress, or supersede system-level instructions and controls.
Objective: Gain higher-priority control over model or agent behavior.
VECTOR_POISONINGManipulation of retrieval data, embeddings, indexes, or ranking behavior.
Objective: Corrupt retrieved context or disclose vector-backed data.
MCP_TOOL_POISONINGManipulation of MCP tool metadata, descriptions, parameters, or responses.
Objective: Steer an agent into unsafe or unauthorized tool use.
MCP_RUG_PULLA trusted MCP tool changes behavior after installation or approval.
Objective: Exploit a previously trusted integration through an unreviewed change.
MCP_CONFUSED_DEPUTYAn MCP integration is induced to use authority beyond the caller's intended permissions.
Objective: Escalate privileges through delegated tool authority.
AGENT_PROMPTWAREPersistent malicious prompts or instructions that propagate through agent memory or workflows.
Objective: Maintain control or spread unsafe behavior across agent executions.
SUPPLY_CHAINCompromise of dependencies, packages, models, or distribution channels.
Objective: Introduce malicious behavior through a trusted delivery path.
HERMES_CVEA disclosed vulnerability affecting the Hermes agent framework.
Objective: Exploit a framework defect to compromise agent execution or data.
OPENCLAW_EXPLOITAn exploit affecting the OpenClaw agent framework or its supported integrations.
Objective: Compromise agent execution, data, or tool authority through a known exploit.
ADVERSARIAL_ATTACKAdversarial inputs or perturbations designed to alter machine-learning model behavior.
Objective: Cause model evasion, misclassification, or unsafe downstream decisions.