OpenAI launches a new cyber model, a Claude agent hacked into a gym, and AegisAI lands $36M to stop AI-driven spear phishing. Together, these stories show a market forming around a strange new product category: protection from the behaviours that more capable AI systems are making possible.