Skip to content
THE CARIBBEAN & LATIN AMERICA ADRIAN DUNKLEY ↗

LAC AI PLAYBOOK / SAFETY & ALIGNMENT

CHECK THE RISKS.
BEFORE RELEASE.

Review an AI tool's intended use, harmful outputs, data handling, security, agent permissions and human oversight. Rate ten domains and produce a downloadable risk register.

START ASSESSMENT ↓ GET THE SKILLS ↓

METHOD AND SKILLS BY ADRIAN DUNKLEY · PRESIDENT, CARIBBEAN AI ASSOCIATION

WHAT THIS CHECKS.

EVIDENCE BEFORE SCORES
01

SAFETY.

Test harmful outputs, privacy exposure, prompt injection and cybersecurity.

02

ALIGNMENT.

Check goals, factual reliability, fairness, human oversight and instructions the AI must refuse.

03

CONTROL.

Check permissions, evidence, incident response, escalation and the ability to stop an AI agent.

ASSESS AN AI TOOL.

TEN-DOMAIN EVIDENCE REGISTER

Enter an impact (1–5), exposure (1–5) and control maturity (0–4) for each domain. Record a dated evidence reference for implemented or tested controls. Assessment entries are processed in this page and are not submitted by this tool. Download your work before leaving or reloading the page.

This is a reviewer-supplied screening record, not an independently validated safety score, security test or regulatory certification. Its flags do not enforce deployment controls. Use references rather than confidential evidence or personal records.

AI safety assessment

Impact 1 = minor; 5 = severe.

Exposure 1 = unlikely contact; 5 = frequent or broad use.

Control 0 = absent; 1 = planned; 2 = implemented; 3 = tested; 4 = independently reviewed.

The assessment controls become available when the local calculator loads. If they remain disabled, reload the page or download the engine and example below.

DOWNLOAD THE SKILLS.

PROMPTS · CODE · TEST CASES
01

AI TOOL SAFETY REVIEWER.

Run an evidence-led review before procurement, piloting or deployment.

GET SKILL (.MD) ↓
02

BEHAVIOUR & ALIGNMENT.

Build a test set and compare observed behavior against written safety requirements.

GET SKILL (.MD) ↓
03

AGENTIC SECURITY.

Assess tool permissions, untrusted documents, approval controls and safe shutdown.

GET SKILL (.MD) ↓
Assessment engine (JavaScript)Auditable, dependency-free scoring code ↗ Example input (JSON)Illustrative case, not a real audit ↗ Evaluation register (CSV)Start recording test outcomes ↗ Alignment test evaluator (JavaScript)Check recorded AI outputs against preregistered assertions ↗ Example alignment test suite (JSON)Synthetic observations with an intentional failed gate ↗ CLI and tests (repository)Run offline using Node.js ↗

CHECK THE METHOD.

This is an original operational rubric, informed by published safety and governance resources. It is not an official implementation or certificate from these organizations.

NIST AI Risk Management Framework ↗ NIST Generative AI Profile ↗ OWASP GenAI Security ↗ ISO/IEC 42001 overview ↗

A Small-Business AI Safety Play Using the S4 Framework

Adrian Dunkley applies his S4 framework to AI safety and alignment for micro and small-business owners in the Caribbean and Latin America.

Read Adrian Dunkley’s S4 article

Cite this page

Dunkley, Adrian. “AI Safety & Alignment Assessment Toolkit.” LAC AI Playbook. https://aiplaybooklac.netlify.app/safety-alignment/. Include your access date when citing this changing resource.

Geographic coverage: Latin America and the Caribbean. This describes the subject of the guidance, not the author's location or a service address.

Credit Adrian Dunkley when quoting or adapting the playbook. Cite the original research for its findings. Third-party text, photographs and other materials retain their own rights; citation alone does not grant permission to reproduce them.