Small models also found the vulnerabilities that Mythos found (aisle.com)

Anthropic’s recent Mythos announcement highlights significant advancements in AI-driven cybersecurity, where its model autonomously identified thousands of vulnerabilities across major software systems. However, a new analysis reveals that smaller, cheaper models also successfully detected many of these vulnerabilities, challenging the notion that only large, frontier models can deliver robust cybersecurity results. Experiments demonstrated that small open-weight models not only uncovered key vulnerabilities showcased by Mythos but often excelled in specific tasks relevant to security analysis, suggesting a jagged landscape for AI capabilities—where performance doesn’t scale linearly with model size. This finding underscores the importance of the overall system and applied security expertise rather than solely relying on the capabilities of high-end models. While Mythos represents a significant contribution to the field, the implications of using a diverse range of less expensive models could reshape the economics of AI in security, allowing broader coverage and facilitating more effective vulnerability detection without the heavy costs associated with deploying massive models. The industry must now focus on integrating these capabilities effectively, ensuring that maintainers can trust the outputs as they transition towards wider application in real-world cybersecurity contexts.