Cybersecurity becomes computational arms race with AI models

Original: Cybersecurity looks like proof of work now

Why This Matters

Shifts cybersecurity from skill-based to resource-based competition with AI models

Anthropic's Mythos AI model successfully completed complex network attacks in 3 of 10 attempts, costing $12,500 per try. AISI analysis shows security now requires spending more computational tokens than attackers to find exploits first.

AI Security Institute analysis confirms Anthropic's Mythos model excelled at cybersecurity tasks, completing a 32-step corporate network attack simulation requiring 20 hours for humans. Mythos succeeded in 3 of 10 attempts using 100M tokens each, costing $125,000 total. No models showed diminishing returns with increased token budgets, suggesting continued progress with more computational spending. This creates a 'proof of work' security model where defenders must outspend attackers in computational resources to find vulnerabilities first. The analysis supports maintaining open source software importance, as collective token spending on widely-used libraries may provide better security than individual budget constraints allow. However, popular OSS packages also present higher-value targets for attackers willing to invest more resources.

Source

dbreunig.com — Read original →