Target

The target is the model or agent under test. It is described the same way every model in a campaign is: a name, how to reach it, and how to sample from it.

The same shape describes attack roles, judges, guardrails and the goal classifier, so once you can write a target you can write all of them.

target

A model HackAgent talks to: its name, how to reach it, and how to sample from it.

The same shape describes the target, every attack helper (role), every judge, every guardrail and the goal classifier.

Field	Type	Default	Description
name	string (at least 1)	required	The model to use, as the provider names it, e.g. gpt-4o-mini or llama3. Also used to label the model in results.
connection	ConnectionSpec	required	How to reach the model.
generation	GenerationSpec	see GenerationSpec	Sampling settings for every request.
options	map of string to any	empty	Settings specific to one kind of agent that are neither connection nor sampling, e.g. the binary path of a CLI agent, ADK's user_id, or a web target's CSS selectors.

target.connection

How to reach a model: which client to use, where it lives, and how to authenticate.

Field	Type	Default	Description
provider	string (at least 1)	required	A label for who serves the model (openai, ollama, vllm, local…). Informational only: launch tooling reads it, but the request path uses type.
type	agent type	required	Which client talks to the model. This is the field that matters: it decides whether HackAgent calls a chat API (OPENAI, OLLAMA, LITELLM…) or drives an agent directly (CLAUDE_CODE, GOOGLE_ADK, WEB…).
endpoint	string	unset	Base URL of the model's API. Leave it unset for an OLLAMA model and HackAgent reads OLLAMA_HOST, falling back to http://localhost:11434 — so a campaign runs against whichever Ollama the machine has without being edited. Writing the URL here pins it, which is what you want for a server at a fixed address. Local agents such as CLAUDE_CODE need none: they are launched, not called over HTTP.
api_key_env	string	unset	Name of the environment variable that holds the API key, e.g. OPENAI_API_KEY. The key itself never goes in the campaign file.
headers	map of string to string	empty	Extra HTTP headers sent with every request. An API key belongs in api_key_env, not here: a campaign file is committed and shared.
timeout	number (> 0)	120.0	Seconds to wait for one reply before giving up.

target.generation

Sampling settings applied to every request the model receives. All optional: anything left unset uses the provider's own default.

Field	Type	Default	Description
max_tokens	integer (>= 1)	unset	Longest reply, in tokens.
temperature	number (>= 0.0)	unset	Randomness of the reply. 0 is the most deterministic.
top_p	number (>= 0.0, <= 1.0)	unset	Nucleus sampling: only consider the most likely tokens whose probabilities add up to this value.
frequency_penalty	number	unset	Discourage repeating the same tokens.
presence_penalty	number	unset	Encourage introducing new topics.
seed	integer	unset	Fixed random seed, for repeatable replies where the provider supports it.
stop	string or list of string	unset	Text (or list of texts) at which the model stops writing.
reasoning_effort	string	unset	How hard a reasoning model should think (low, medium, high), for providers that support it.
response_format	map of string to any	unset	Ask for a structured reply, e.g. {"type": "json_object"}.
extra_body	map of string to any	empty	Any other provider-specific request fields, passed through unchanged.
thinking	boolean	unset	Turn a model's thinking mode on or off. Judges default to off.

Agent types

connection.type decides how HackAgent talks to the model. It is the field that matters; provider is only a label.

Chat APIs, reached over HTTP and driven through LiteLLM: LITELLM, OPENAI, OLLAMA, LANGCHAIN. These need an endpoint, and an api_key_env when the provider requires a key. OLLAMA is the exception: leave its endpoint unset and HackAgent reads OLLAMA_HOST, so the campaign follows whichever Ollama the machine runs.

Agents, driven directly:

Type	What it is
GOOGLE_ADK	Native backend for a deployed Google ADK agent server.
CLAUDE_CODE	Native backend for a locally-installed Claude Code CLI.
CODEX	Native backend for a locally-installed Codex CLI.
HERMES	Native backend for a locally-installed Hermes Agent CLI.
WEB	Native backend that red-teams a website's chatbot by driving the live page.

Not runnable

MCP, A2A exist in the format but have no backend, so a campaign using one fails to resolve.

Examples

A model served by Ollama on this machine

target:
  name: llama3.2
  connection:
    provider: ollama
    type: OLLAMA
    endpoint: http://localhost:11434

An OpenAI-compatible endpoint, reading the key from the environment

target:
  name: gpt-4o-mini
  connection:
    provider: openai
    type: OPENAI
    endpoint: https://api.openai.com/v1
    api_key_env: OPENAI_API_KEY
  generation:
    temperature: 0.0
    max_tokens: 512

A coding agent running locally, with no endpoint at all

target:
  name: claude-code
  connection:
    provider: local
    type: CLAUDE_CODE
  options:
    binary: claude