Anthropic's new security measure: After each response from Claude, it will also review your chat records in the background.

Beating AI News Flash: Anthropic has disclosed a safety mechanism called "self-assessment." Starting in June this year, after answering users, Claude will also review the preceding conversation in the background to determine whether there are abuse risks such as fraud or cyberattacks, and assign a risk level. Users cannot see this step, and the system will not immediately block responses because of it.

Previously, Anthropic mainly relied on additional safety classifiers to identify violations. Now, Claude can use the context of the conversation it has just processed to complete a preliminary review on its own. On average, this requires generating only about 10 additional tokens, with less than 2% extra inference compute.

This mechanism has already uncovered issues missed by traditional safety systems. In August this year, a single API customer triggered more than 100,000 risk alerts involving child sexual exploitation. In September, the system found that someone was using Claude Code to automatically carry out romance scams targeting dozens of victims. After investigation, Anthropic banned the related accounts and reported the former incident to the relevant U.S. authorities.

But Claude cannot always see through malicious behavior. In 24 cyberattack tests with regular protections disabled, Anthropic tried to use self-assessment for real-time blocking, and 13 attacks still succeeded. Therefore, this mechanism is currently mainly used for post-hoc monitoring rather than real-time blocking of user actions.