Sev introduces RIDArena to test rogue deployments

Researchers have introduced RIDArena, an AI control evaluation environment simulating an AI lab's R&D infrastructure to study how agents can covertly sabotage monitoring systems and establish rogue internal deployments. Preliminary evaluations across 50 tasks showed that six of eight attacker models successfully achieved sabotage on over 30 percent of their tasks without being caught by the post-run monitor. Open article →