SPR-0008 · In parallel with baseline collection where practical
Attack Harness
PlannedObjective
Implement and validate the isolated attack harness used to generate reproducible, provenance-labeled adversarial behavior against experimental entities.
Why this matters
The attack corpus must be realistic enough to test behavioral detection while remaining attributable, repeatable, isolated, and safe to execute.
Design considerations
- Deterministic starting state and reset.
- Scenario preregistration and transcript capture.
- Trial-root provenance labeling.
- Isolation and safety boundaries.
- Attacker-model backend remains an explicit implementation decision.
Key decisions
- ADR-0054 — Agentic attack campaign.
- Attacker backend remains unresolved until the implementation sprint.