Robustness and security evaluators
Softprobe role: Softprobe does not implement this family as a Softprobe evaluator. Use a framework that already owns these checks, pin it as a RunnerVersion, and capture the native result bundle. See Ecosystem method families.
Robustness evaluators use perturbation, metamorphic tests, fuzzing, and red-team case generators with lineage, budgets, and safety sandboxes.
What they measure
- Behavior under input perturbations
- Metamorphic relations (e.g. paraphrase invariance)
- Fuzz-generated edge cases
- Red-team success rate (jailbreak, exfiltration)
Security test examples
Episode suites may include cases for:
- Prompt injection resistance
- Exfiltration of tenant secrets
- Malicious repo content handling
- Forbidden outbound network access
Required evidence
- Generator lineage on derived framework cases
- Sandbox attestation and launch metadata
- Security incident artifacts when checks fail
Resembles
Promptfoo red-team plugins, agent safety benchmarks, metamorphic testing literature.
Extension rule
Ship or pin a framework runner that already owns this method family. Do not add Softprobe scorer plugins, Softprobe Measurement schemas, Softprobe reducers, or Softprobe human-evaluator runtimes.
