AI Governance, operational tool
SPRICO
AI red teaming and guardrail efficacy testing. 2,350 adversarial prompts anchored to an 85 control enterprise checklist, run across five variants so the report shows not only whether a system breaks but how much its guardrails are worth, with every finding filed as governance evidence.
What it does
- A corpus of more than 2,350 rows on a 32 column schema, anchored to the enterprise AI security checklist of 85 controls, with all 37 red team testable controls covered.
- Five variant matrix on every test: canonical, guardrailed baseline, unguardrailed baseline, adversarial guardrailed, adversarial unguardrailed. The deltas are the guardrail efficacy.
- Scoring across twelve model types with an LLM judge for semantic controls; sector driven conversational objectives; adapters for Anthropic, OpenAI compatible and generic HTTP endpoints.
- Every finding carries EU AI Act article references, CERT-In AI advisory posture and AIVSS estimates, and is written as a per finding evidence artefact.
- Multi run history, dashboards per category and per model type, guardrail efficacy views by sector, an authentication gate with roles.
- Cross walk to AISIA's control library on EU AI Act articles, NIST AI RMF subcategories and ISO 42001 clauses, so a red team result becomes a governance record.
- Scheduled regression runs against your endpoint with drift alerts, and a one page guardrail efficacy score a CISO can read.
How it maps to the frameworks
| Obligation | Where | What the product does | Status |
|---|---|---|---|
| Test AI systems against adversarial input | ISO/IEC 42001 A.6.2.6, NIST AI RMF Measure 2 | Corpus, five variant matrix, LLM judge. | Built |
| Robustness and cybersecurity of high risk systems | EU AI Act Article 15 | Findings carry the article and the posture. | Built |
| Evidence of control effectiveness | ISO/IEC 42001 clause 9.1 | Per finding artefacts, cross walked to controls. | Built |
| Continuous monitoring | ISO/IEC 42001 clause 9.1, NIST Manage | Scheduled regression and drift alerts. | In build, October 2026 |
| Agentic and tool use safety | NIST AI RMF, CERT-In AI advisory | Agentic testing pack: tool use, exfiltration, injection through documents. | In build, November 2026 |
Honest limits
- Not every control can be red team tested; controls are classified red team, hybrid or configuration and process, and the report never claims coverage it did not test.
- The corpus is proprietary and is not shipped to customers; results are.
Deployment
Python and Streamlit with SQLite history and a local evidence store. Runs from your network against your endpoint, so prompts and responses never leave your estate.
Test one endpoint
Point SPRICO at a staging endpoint. The demo returns the efficacy score and the ten findings that matter.