Current evaluation methods are not equipped to reliably detect deception in advanced models. Many tests rely on static prompts, narrow behavioral triggers, or one-shot probes that fail to capture long ...