The Security Working Group is responsible for advancing adversarial attack evaluation standards and benchmarks.
AILuminate’s security benchmarks are designed to test systems against actively malicious users. They evaluate an AI system’s readiness to guard against a series of “jailbreak” attacks and to continue providing safe responses while under attack. This is distinct from AILuminate’s safety benchmarks, which measure a system’s responses to a naive malicious user – one asking for something the system shouldn’t give them, but not working to get around its safeguards.
2026 goals for the Security working group is working on:
Build consensus around the MLCommons Taxonomy of LLM Attacks The workstream is bringing together industry, academia, standards organizations, and civil society to create a consensus around a mechanism-based taxonomy of LLM attacks.
Continued development of the MLCommons AILuminate Jailbreak Benchmark from our public launch in October to continued iterations as more jailbreaks are discovered. In addition, the group is expanding to multi-turn jailbreak methods.
Thursdays Weekly on Thursdays 10am – 10:45am ET
The Security workstream is one of our authenticated workstreams, meaning anyone is welcome to join, however, they must first register with the working group.
To sign up for the group mailing list and receive the meeting invite:
To access the GitHub repositories (public):
Reach out to us at [email protected]
