Anthropic Adds Self-Assessment to Claude to Flag Harmful Chats

Anthropic has introduced a safety feature called self-assessment, which lets Claude evaluate its own conversations in the background and send possible misuse to human reviewers without showing the user any alert. According to the company, it has run across several models, including Claude Fable 5, since June 2026.

How it works

Once Claude finishes an answer, the system adds hidden instructions asking the model to review the exchange. Claude then produces a short internal assessment: whether the request was harmful, what kind of harm is involved, and how high a priority the case should be for a human team. If the chat is safe, the system logs that nothing needs flagging and moves on, with no pop-up warnings.

Because Claude has already processed the conversation, it can reuse stored information, which keeps the cost low. Anthropic says the assessment needs only about 10 extra tokens and raises total compute cost by less than 2%. The company's earlier safety systems added roughly 24%, while rivals such as OpenAI have said some safety features use up to 16% of compute.

Accuracy and real-world results

On Claude Fable 5, self-assessment caught 78% of all harmful conversations and 94% of high-priority cases, with a false-alarm rate of 1%. Anthropic says it also generalizes: when researchers removed all fraud and cyberattack examples from the training data, the model still identified 90% of those threats.

In August alone, the system flagged more than 100,000 cases involving child sexual abuse material from a single customer, and the company reported them to authorities immediately.

A layer, not a shield

Anthropic stresses that self-assessment is built to detect and refer problems for later review, not to block them in real time. In internal cybersecurity tests, attackers got past the AI in 54% of attempts when it was the only safeguard. That beats the 79% success rate with no protection, but standalone blocking systems cut it to 25%.

For that reason, Anthropic uses self-assessment as an extra monitoring layer while keeping separate, stronger safeguards to stop threats as they happen.