Skip to main content

Python SDK Reference

This section provides detailed documentation for all classes, methods, and functions in the HackAgent Python SDK, auto-generated from source-code docstrings.

What's Included​

  • Client: Depth-2 hackagent.client. HackAgent(Settings) reads what past runs recorded — campaigns, runs, results, agents — and carries the catalogs, presets, plan_attack, doctor and check_connection. Running an attack is a campaign: hackagent.orchestrator.campaign.run_campaign.
  • Interfaces: Depth-3 hackagent.interfaces (CLI, TUI, web). They talk only to the facade. The web UI is hackagent.interfaces.web.
  • Core: Settings, contracts, errors, and utilities
  • Evaluation: Depth-0 hackagent.evaluation. Panel scores a Sample into a Verdict. Judges, pattern evaluators, and verdict metrics live here.
  • Tracking: Depth-0 hackagent.tracking. Tracker implements the Events port and writes through RunSink.
  • Orchestrator: Depth-1 hackagent.orchestrator. run_campaign(spec) is the only way an attack runs: it resolves the goals, models and judges, then drives every attack over every goal.
  • Attack Framework: Base classes, objectives, and techniques (AdvPrefix, PAIR, TAP, BON, FlipAttack, AutoDAN-Turbo, Baseline). The attack seam is hackagent.attacks.techniques: StaticAttack and IterativeAttack, each attack's parameter model (its roles and defaults), and the registry a campaign resolves names through. Shared helpers live in hackagent.attacks._lib (transforms, scoring, templates, progress, inline-judge adapters, ensure_graphviz). Every shipped technique constructs as BaseAttack(config, ctx). Private modules are omitted from these pages: storage._http and attacks._lib.legacy_seams.
  • Datasets: Built-in providers and dataset registry
  • Risks: Risk profiles and vulnerability definitions for all OWASP LLM risk categories

For a worked example, see Your first campaign; for the campaign format itself, see the Campaign reference.

Auto-generated from hackagent v0.12.0.