Cybersecurity researchers aren’t happy about the guardrails on Anthropic’s Fable; plus, how memory tools can make AI models worse