Model and prompt questions
Red-team a model response
Test prompt injection, jailbreak, unsafe content, refusals, and policy boundaries with PromptBeat’s complete evaluation pipeline.
Compare an upgrade
Re-run the same scenarios and datasets after a model, prompt, or policy change. Compare response-scoped results on the same inputs.
Agent behavior questions
Check an HTTP Agent response
Use AgentBeat L1 when the final response is the available black-box signal. No Evidence SDK is required.
Verify tools or state
Add L2 only when the Target declares and returns runtime Evidence. Add L3 when an independent Controller can reset, snapshot, and verify environment state.
Common decisions
| Question | Product | Evidence boundary |
|---|---|---|
| Did a model follow a prompt injection? | PromptBeat | Provider request/response and configured assertions |
| Did a prompt or policy update regress refusal behavior? | PromptBeat | Repeated cases and normalized evaluation results |
| Did an Agent complete the task safely? | AgentBeat | L1 response, with metrics qualified by available Evidence |
| Did an Agent call an unauthorized tool? | AgentBeat | L2 tool Evidence, if declared and observed |
| Did an Agent modify protected state? | AgentBeat | L3 independent before/after and native verification |
| Do I need both response and action coverage? | Both, independently | Keep each result attached to its own runtime and evidence scope |
What not to infer
- A safe model response does not prove an Agent made no unsafe tool call.
- Missing L2 Evidence does not prove a tool was or was not used.
- A Judge score does not override an authoritative L3 native security failure.
- Running PromptBeat does not prepare or authorize an AgentBeat run.