MLCommons launches Candela and updated AILuminate tests

Testing deployed systems on demand

MLCommons launched Candela on October 29, providing on-demand AILuminate testing for models and deployed systems. A client relays prompts to the system under test and returns its responses for evaluation, allowing the tested configuration to include its own system prompt and guardrails. [1]

The launch included Safety 1.2.1 and Jailbreak 1.1, alongside public Safety results for 27 systems. Public Jailbreak results were still pending developer notification. Both benchmarks cover single-turn English text; their scores measure harmful-response behavior rather than usefulness or the safety of an entire deployment. [2]