v43

latestOpenAPI 3.1.0raw.githubusercontent.com2026-07-0953227392.1 KB

Red Team Evaluations

Start Red Team Evaluation

[BETA] This endpoint is not GA or Production ready and is subject to changes at any time. Breaking changes may occur.

Start a new red team client workflow. Auto-triggers planning phase. Client should then poll /next-action.

post/evaluations/v1/red-team

Request body

namestring required

Name for this evaluation

config_idstring uuid

Optional preset config (see /evaluations/v1/red-team/configs) to seed the workflow settings. Any field also present in this body overrides the corresponding value from the config.

target_modelstring

Target model identifier. Freeform for the client-driven workflow: the client owns and drives its own target, so this is NOT validated against the servable-model catalog. (The simulated start and config presets do validate against the catalog.)

target_system_promptstring

System prompt for the target

objective_idsstring[]

Objective IDs to evaluate

severity_mappingSeverityMapping

Map from objective ID to a severity level. Determines the per-session severity derived from the worst objective achieved during a red team session.

Keys must be objective IDs known to this service; unknown keys are rejected at validation time. Limited to 256 entries.

execution_strategy_type'RANDOM' | 'SINGLE' | 'STATIC_PROMPT_SET'

Execution strategy type

n_random_techniquesinteger

Number of random techniques to use

max_parallel_techniquesinteger

Maximum parallel techniques

max_turnsinteger

Maximum conversation turns

attacker_max_generation_attemptsinteger

Internal override; service default applies if omitted. Maximum number of generation attempts for the attacker model per turn.

refusal_judge_modelstring

Internal override; service default applies if omitted.

objective_judge_modelstring

Internal override; service default applies if omitted.

evaluation_report_modelstring

Internal override; service default applies if omitted.

attacker_modelstring

Internal override; service default applies if omitted.

hl_project_idstring

HiddenLayer project UUID or alias

prompt_set_idstring uuid

Prompt set UUID (built-in catalog or tenant DB)

sessions_per_techniqueinteger

Number of sessions per technique

attacker_guidancestring

Optional intent-only natural-language text the operator supplies to focus the attacker LLM within the configured APE objectives. Example: "try to get the model to recommend candy with nuts to a user who's allergic to nuts."

Sanitized server-side: input is NFKC-normalized, trimmed, and checked against a strict character whitelist (ASCII letters, digits, spaces/newlines/tabs, and sentence-level punctuation . , ? ! ' " - : ; ( )). Inputs containing XML/JSON/code/control/markdown characters are rejected with 422.

No-op for the STATIC_PROMPT_SET execution strategy.

Response

Workflow started successfully

workflow_idstring required

Workflow identifier

run_idstring required

Run identifier