Anthropic’s Opus 5 Nears Mythos 5 on Finding Bugs, but Falls Short on Exploits
Source SecurityWeek
Description
Anthropic released Opus 5, an AI model capable of identifying software vulnerabilities nearly as well as their top-tier Mythos 5 model but with significantly reduced capability in generating exploits. Opus 5 blocks binary-based vulnerability scanning, penetration testing, and exploit generation to limit offensive cyber use. These restrictions are deliberate, reflecting Anthropic's design choices to avoid training Opus 5 on offensive tasks. The model's safety classifiers reduce interventions compared to previous versions, but requests triggering offensive capabilities fall back to an older model. Anthropic offers a version with fewer restrictions to select enterprises and researchers. Mythos 5 remains withheld from general release. The models were temporarily taken offline due to U.S. government concerns but later reinstated.