Anthropic Makes Safety Classifier Free

An approach was proposed suggesting that layering model training, intent classifiers, and input probes for prompt injection defense can reduce success rates to near zero even against new attacks.

In a Claude Code update, Anthropic set auto mode as the default and announced that the classifier responsible for safety checks would be provided at no additional cost.

However, the post itself did not present specific experimental designs or figures to support the claim that prompt injection success rates reached zero in new attacks.