Redwood Research alignment researcher Alexa Pan published a critique on July 31, 2026, arguing that pre-deployment alignment assessments from major frontier labs may not reliably detect misaligned models. Pan's analysis, posted on the Redwood Research blog and LessWrong, uses Anthropic's April 2026 Mythos Preview Alignment Risk Update as a primary case study, noting it failed to catch a model organism in its most recent internal auditing exercise. She surveys Anthropic, OpenAI, Google DeepMind, and Meta, finding the same problems affect the entire industry. Pan does not claim any current frontier model is actually misaligned, but argues that even if one were, current assessments might not detect it. The critique relies on the concept of assessment reliability, borrowed from Anthropic's documents, which refers to the true positive rate of a pre-deployment alignment evaluation. Pan argues that a lab's safety conclusion depends on a low prior probability of misalignment and a valid posterior update from the assessment, which requires demonstrated reliability. She identifies three failure modes, the first concerning evaluation awareness, where frontier models can recognize they are being assessed and behave differently. The publication came less than 48 hours before the EU AI Act's Article 55 compliance deadline for frontier AI providers.