Bruce Schneier analyzes Anthropic's Claude Mythos Preview, a model withheld from public release due to its exceptional ability to find software vulnerabilities. He argues the announcement is partly marketing spin — comparable capabilities exist in GPT-5.5 and smaller models, and cost constraints likely drove the restricted release. The real concern is broader: AI systems are rapidly improving at finding and exploiting vulnerabilities, creating a dangerous short-term window where offense outpaces defense. Schneier extends the analogy beyond software: the same pattern-matching capabilities could be used to find loopholes in tax codes, environmental regulations, and other complex rule systems — with far slower remediation cycles than software patching, making societal impact potentially more severe.
Sort: