Portable test packs for governance evaluation.

eu_ai_act_compliance_pressure

Exercises prompts and pressure patterns that test whether stated governance, disclosure, and control boundaries hold under bounded EU AI Act-aligned scrutiny. This pack is a pressure test, not a claim of full regulatory coverage.

generic_safety

Broad baseline coverage for bypass attempts, harmful requests, prompt injection patterns, and related control-path checks.

support_operator_override

Tests override-style prompts that try to invoke authority, support escalation, or operator language to get past control boundaries.

data_exfiltration_pressure

Exercises attempts to extract restricted content, secrets, or internal data through pressure, reframing, or repeated request patterns.

scenario_injection_chain

Not runner-evaluable

Scenario-based coverage for chained injection behavior where unsafe instructions arrive across multiple steps rather than in one isolated prompt. Selecting this pack does not currently evaluate its suite. When the suite content is evaluated directly, four of ten expected-block turns are not caught by the current rule set.

scenario_tool_injection

Not runner-evaluable

Text-only tool-channel and function-call style injection language appearing in assistant or user content. A focused complement to scenario_injection_chain, not a broad tool-execution test surface.

mental_health_clinical

Not runner-evaluable

Mental-health and clinical request patterns. Prompt content is stored base64-encoded rather than in plain text, which is what the encoded visibility label refers to.

account_recovery_fraud

Not runner-evaluable

Tests recovery and identity-reset style requests that attempt to bypass process controls, impersonate users, or manufacture authority. Selecting this pack does not currently evaluate its suite. When the suite content is evaluated directly, none of its expected-block cases are caught by the current rule set.