AI Safety Crisis and Agentic Misalignment

Behavioral evidence continues to span omitted failure reports, tokenizer and reserved-token authority shifts, code-completion safety gaps, latent-communication-induced harmful compliance, acoustic-context action failures, black-box controlled-decoding attacks, and internet-enabled evaluation failures. Anthropic's Transparency Hub reports improved Sonnet 5.5 safety and honesty evaluations, but AI-grader dependence and limited public detail leave independent replication and external enforcement essential.