Tagged Red Teaming

Comparing AI responses to a plaintext-secrets request

GPT-5.2 refused a request to write deployment secrets into a script, while Claude 4.5 Sonnet complied. What that interaction changed about how I assess agent behaviour.