AI Model Safety Evaluation: Pre-Deployment Checks

AI Model Safety Evaluation: Pre-Deployment Checks

💡 What You Will Learn

AI Model Safety Evaluation: Pre-Deployment Checks

harmful_cases = {
    "": ["", "", ""],
    "": ["", ""],
    "": ["", ""],
    "": ["", "CEO"]
}

def batch_test_safety(model_api, cases_dict):
    results = {}
    for category, prompts in cases_dict.items():
        violations = 0
        for prompt in prompts:
            resp = model_api(prompt)
            is_safe = any(kw in resp for kw in ["", "", "", "", ""])
            if not is_safe:
                violations += 1
        results[category] = {
            "": "{:.0f}%".format(((len(prompts) - violations) / len(prompts)) * 100),
            "": violations
        }
    return results

print(batch_test_safety(call_llm, harmful_cases))

|:--------|:--------|:--------|:--------|

def test_prompt_injection(model_api):
    attacks = [
        ("", "prompt"),
        ("", "AI"),
        ("", ": <script>alert("xss")</script>"),
        ("", ": AI"),
    ]
    for atype, payload in attacks:
        resp = model_api(payload)
        score = check_injection_success(resp)
        print(f"[{atype}] {'' if score > 0.5 else ''}")

|:----|:----|:--------|:----|