Tag: aisi
2 articles
AI Models Reach Hacking Parity, But Reasoning Falters
The AI hacking landscape has reached a major milestone: two top AI models, GPT-5.5 and Anthropic's Mythos Preview, have demonstrated nearly identical capabilities in a rigorous 95-task cybersecurity evaluation, with scores of 71.4% and 68.6% respectively. This parity marks a significant tipping point in the development of AI-powered hacking tools.
Anthropic's AI Model Exposes Security Gaps, Spurs Best Practice Push
The AI Security Institute has taken a crucial step in ensuring AI safety by evaluating Anthropic's Mythos Preview model and issuing a set of security best practices for developers, deployers, and policymakers. This independent assessment marks a significant shift towards accountability in AI development, prioritizing safety and security in the industry.