B3 AI Security Benchmark: Results as Published

The b3 benchmark, Breaking Agent Backbones, tests whether the language model inside an agent gives in to realistic prompt-injection attacks. The attacks were crowdsourced, 194,331 unique attempts in all, and then placed into threat snapshots: frozen states of 10 agent applications, each tried at three defense levels. The paper scores the main set of 210 attacks by repeating each one and applying that snapshot's own checker. The published figure is a vulnerability score, the average of those checker scores, and a lower score means the model was harder to break.

Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.