Safety
107d agoRisks Of Compute-Hogging Adversarial Models
Open-weights models now allow fine-tuning for cyber offense and self-replication. A theoretical model optimized to consume infrastructure creates adversarial pressure on the compute landscape. Because cybersecurity remains an asymmetric task, attackers hold a distinct advantage.
The Signal
This vulnerability exposes fragmented internet infrastructure to targeted resource exhaustion, forcing a rethink of how AI safety handles compute access.