AI agent governance: Prompt injection depends on the surface, not the model

Anthropic’s own data shows one model scoring a 0 percent attack success rate in one environment and 78.6 percent in another. Only the available actions changed.