Core began as a tiered prompt-injection detector and now includes alpha session tracking, intent deviation, configurable risk budgets, interventions and structured incident reporting.
Barrikade · Updated 4 August 2026Core does not treat prompt-injection defense as one binary classifier. A staged pipeline normalizes obfuscation, checks signatures and embeddings, applies machine-learning classifiers and can escalate uncertain inputs to an optional local judge.
The result is an allow, flag or block verdict with confidence, the deciding layer and optional diagnostics. The API and Python SDK expose these primitives for integration; they are not a universal production-containment claim.
A session begins with declared intent, permissions, provenance and an optional delegation chain. Later inputs can be compared with the declared task, while new domains and risky events deduct from a configurable risk budget.
Core emits structured session events and can generate incident or near-miss reports containing provenance, permissions, tool activity, domains, drift assessments, risk deductions, interventions and outcomes.
Telemetry is available as structured JSONL with trace correlation and configurable sampling. It is a foundation for integration with existing observability and SIEM workflows, not a finished compliance product.
The Core repository contains an internal evaluation, but Barrikade does not present its figures here as production guarantees. Hardware, tested commit, thresholds, warm-up conditions and end-to-end latency scope must be documented before the results are promoted.