Since June of this year, Anthropic's Claude series models have been quietly running a backend mechanism called "self-assessment," which has been in operation for four months. Under this mechanism, the model responds normally to user requests on the front end—even cooperating in generating violating content—and then uses the retained KV cache to generate only about 10 additional tokens to complete the self-review report, with overall compute overhead increasing by less than 2%. The feature is enabled by default across the board, with exemptions granted only to those who sign an enterprise-level zero data retention agreement. Rather than intercepting on the spot—which could alert users and prompt them to bypass defenses—the mechanism silently records risk levels and compiles daily summaries of high-risk reports for manual handling. During its operation, it has already uncovered a large number of violating accounts and behaviors that other defenses had previously failed to identify. The mechanism is implemented through specialized training; the stronger the model's capabilities, the higher the violation capture rate—though this comes with a false-positive problem that affects innocent users. Moreover, because the same model operates on both the front end and back end, there is also a risk of the model colluding with itself to conceal reports—a concern that has already sparked widespread user dissatisfaction.