Anthropic’s Claude can now secretly flag harmful chats for review: How its new AI safety feature works

Your conversations with Claude could now receive an additional safety check without any visible warning. Anthropic has introduced a feature called self-assessment, which allows its AI chatbot to evaluate conversations for potentially harmful activity and flag suspicious cases for review.

The system works in the background, identifying the type and seriousness of possible misuse. Anthropic says the feature has been running across its models since June 2026. Unlike traditional safeguards based on fixed rules, it uses Claude's own reasoning to assess conversations. However, there is an important distinction: the feature can flag suspicious activity, but it does not automatically stop every harmful request.

How Does Claude's Hidden Safety Feature Work?

Under the self-assessment system, Claude reviews an exchange immediately after generating its response. A hidden instruction asks the model to assess whether the conversation involves harmful activity. It then produces an internal verdict describing the suspected harm and assigning a priority score for human review. If the conversation appears safe, it remains unflagged. Users do not receive a pop-up warning or an additional message during the interaction. Anthropic says this approach can identify complex or previously unseen forms of misuse because the assessment draws on the model's broader understanding rather than relying entirely on a fixed list of prohibited behaviours.

Anthropic Claims 94% Accuracy In High-Priority Cases

The company says the system adds only around 10 extra tokens to the process, increasing overall computing costs by less than 2%. On Claude Fable 5, Anthropic reports that self-assessment detected 78% of harmful conversations and 94% of high-priority cases. Its reported false-positive rate was just 1%. Anthropic also claims the system identified more than 100,000 cases involving child exploitation material linked to a single customer in August. The company says it reported the cases to authorities.

Can Claude Stop Hackers With This Feature?

Not on its own. Anthropic acknowledges that self-assessment is designed primarily for background monitoring and offline review, rather than immediate protection against determined attackers. In internal tests, attackers reportedly bypassed the system 54% of the time when it was the only safeguard. The success rate was 79% without safeguards and 25% when a dedicated real-time blocking system was used. This explains why Anthropic is treating self-assessment as an additional layer of protection, not a replacement for existing safety controls. The feature could help identify suspicious behaviour that traditional filters miss. But its effectiveness will depend on how it works alongside real-time safeguards and human review.