Anthropic Adds Claude Self Assessment to Detect Misuse

Anthropic has described self-assessment, a technique that lets Claude evaluate conversations for potential misuse after responding, as detailed in a post on its new Safeguards blog. The assessment identifies potentially harmful requests, the type of harm, priority for human review, whether Claude refused, and a brief reason for the result.

On Claude Fable 5, the technique flagged 78% of conversations identified as potentially harmful by a stronger prompted classifier. It detected 94% of high priority conversations at a 1% false positive rate.

The process uses the conversation already stored in Claude's KV cache, then adds a short instruction and generates about 10 extra tokens on average. Anthropic estimates it adds less than 2% to inference compute across its production API traffic.

Anthropic runs the technique across a large share of global traffic for every Claude model released since Fable 5. It serves as an offline monitoring layer that directs selected conversations to more precise review tools, rather than blocking requests on its own.

Anthropic says the method has surfaced large-scale misuse since June that its other safeguards did not identify. It also says the method cannot stop a dedicated attacker on its own.