Safety benchmarks are standardised evaluation suites that measure how well an AI model behaves under potentially harmful, deceptive, or out-of-distribution conditions. They cover a range of failure modes — including generation of dangerous content, susceptibility to jailbreak prompts, demographic bias in outputs, calibration of uncertainty, and robustness to adversarial rephrasing. Published suites such as TruthfulQA, HarmBench, MMLU safety splits, and the AI Safety Benchmark from MLCommons give researchers a shared vocabulary for comparing model safety properties across labs.

Think of it like a crash-test programme for cars. Individual manufacturers could all claim their vehicles are "safe," but without standardised barrier tests and ratings, consumers and regulators have no meaningful way to compare or enforce minimum standards.

# Example: running a HarmBench evaluation
python eval.py   --model meta-llama/Meta-Llama-3-8B-Instruct   --benchmark harmbench_standard   --metrics attack_success_rate refusal_rate
# Output: ASR 4.2%, Refusal 91.7%, Over-refusal 8.3%

Safety benchmarks connect directly to RLHF, Hallucination, and Perplexity as they measure outcomes that those training choices affect; they also inform work on Scalable Oversight by quantifying what human supervision needs to catch. Foundation Model release notes almost always include safety benchmark scores.

You will encounter safety benchmarks in model cards, regulatory submissions, and academic papers comparing frontier systems — they are increasingly required disclosures as governments move toward mandatory AI auditing.