What’s next for SCAM and agent trust
These results are preliminary. The benchmark is new, and we're sharing it early because we think the findings are important enough to warrant public discussion even before we cut a 1.0 release. There's a lot of work left to do, and we'd love your help. We need more scenarios covering new attack types, more models on the leaderboard, better scoring rubrics, and stress-testing of the benchmark itself. If you're a security researcher, an AI engineer, or someone who builds agents that handle credent
13 juil.