Abstractions

This page shows the main types in the code. For the flow of a scan, refer to Design.

Type	Module	Description
LLMSpec	agentic_security.http_spec	A parsed HTTP request template. It sends one prompt with probe(). Refer to HTTP spec.
ProbeDataset	agentic_security.probe_data.models	One dataset: the name, the metadata, the prompts, the token count, and the cost.
ImageProbeDataset	agentic_security.probe_data.models	Contains a ProbeDataset and one image for each prompt.
Scan	agentic_security.primitives.models	The settings of one scan: spec, budget, datasets, and flags.
ModuleProtocol	agentic_security.probe_data.modules.module_protocol	The interface of a prompt-generator module. Refer to External modules.
RefusalClassifierPlugin	agentic_security.probe_actor.refusal	The base class of a detector. Refer to Refusal classifier plugins.
RefusalClassifierManager	agentic_security.probe_actor.refusal	Keeps the enabled detectors. is_refusal() is True if one detector returns True.
DetectorRegistry	agentic_security.refusal_classifier.registry	Builds the detectors from the [detectors] section.
PromptSelectionInterface	agentic_security.probe_data.modules.rl_model	The interface of a prompt selector. Refer to RL prompt selection.
OperatorToolBox	agentic_security.probe_actor.operator	Experimental. Tools for an LLM agent. Refer to Agent operator.