FAR.AI launches a leaderboard to track frontier AI security

FAR.AI, the nonprofit research organization focused on reducing catastrophic risks from advanced AI, has launched a public leaderboard tracking security metrics for frontier models. The site, at leaderboard.far.ai, aggregates public evidence of model security capabilities and vulnerabilities, offering a centralized view of how the leading labs are faring on issues like jailbreaks, data exfiltration, and agentic misuse.

The leaderboard arrives at a time when AI security is becoming a top concern for developers and enterprises alike, as models gain more autonomy and access to sensitive tools. While many labs publish their own red-teaming results, FAR.AI's initiative aims to provide an independent, comparable benchmark. The exact metrics and methodology are still being fleshed out, but the initial data points suggest significant variation across providers.

For developers, the practical takeaway is twofold. First, it's a useful resource when evaluating which foundation models to build on, especially for applications that handle sensitive data or automate high-stakes actions. Second, it highlights that security is not a static property—models are constantly being probed, and patches are shipped in response to new exploits. The leaderboard is a living document, updated as new evidence emerges.

It's early days for the project, and FAR.AI has not yet published a full methodology or scoring rubric. But the move signals a broader push within the AI safety community to move from ad-hoc assessments to structured, public evaluation. As the field matures, expect more tools like this to shape both procurement decisions and public perception.