Anthropic's new AI is learning to hack. The security industry is about to get automated out of a job.

Anthropic released new Claude AI models, Fable 5.2 and Mythos 5.2, boasting improvements in coding, reasoning, and vision. The company claims they helped engineers with complex tasks like software migrations and even formalized proofs for complex math theorems. More notably, they are being benchmarked on their ability to plan and execute multi-stage cyberattacks in simulated environments.

In a key benchmark called CyScenarioBench, Fable 5.2 completed over 76% of challenges, outperforming its predecessor and OpenAI's specialized 'GPT-6 Cyber' model. It also scored nearly 90% on an internal test designed to see if it could rediscover known software vulnerabilities from source code alone. The takeaway is clear: AI isn't just an assistant for defenders anymore; it's rapidly becoming a sophisticated tool for automated offense, turning cybersecurity into an AI vs. AI arms race.