← All questions
Security & SafetyHard· Data Poisoning· Evaluation· Red Teaming

Your team shipped a per-source vote cap as the poisonin...

A per-source vote cap only protects against a single source flooding the system with near-duplicates; it does not account for coordinated, multi-source attacks where each source stays under the individual threshold. The fact that the benchmark passed is irrelevant if it did not model the specific threat of a coordinated adversary. To secure the system, you need additional defenses that can detect patterns of coordination across multiple distinct sources.

Code playground (Python)