Claude can now flag your chats to Anthropic if it detects indecent behaviour

Anthropic has rolled out a new safety feature for its AI model, Claude, allowing the chatbot to silently evaluate its own conversations for potential misuse, called "self-assessment." This background monitor has been running across models like Claude Fable 5 since June 2026.Taking to X, Anthropic shared that the technique trains Claude to give a preliminary safety assessment right after it responds. Because the tool draws on Claude’s own judgment rather than a fixed list of rules, it gets more accurate as the underlying AI models improve and catches complex misuse that older safeguards were never designed to spot.How Claude self-evaluates behind the scenesTypically, interacting with an AI involves a simple back-and-forth exchange. With self-assessment, an extra step happens completely out of sight. Right after Claude generates an answer, the system adds a hidden instruction asking the AI to review the exchange. Claude then produces a short internal verdict, noting if the request was harmful, the type of harm, and a priority score for human review.If the conversation is safe, the system simply registers it as unflagged and moves on. This happens without interrupting the user or generating annoying pop-up warnings. Because Claude has already read the conversation, it can reuse the stored memory of the chat, making this extra step highly efficient. High accuracy without the heavy computing costSafety checks usually require a massive amount of computing power, but Anthropic’s new method is surprisingly cheap. The process takes just about 10 extra tokens, or small chunks of text, to run a check. This adds less than 2% to the total computing cost.To put that in perspective, Anthropic’s older safety systems added nearly 24% to costs, while rivals like OpenAI have reported safety features consuming up to 16% of computing power. Despite being lightweight, the tool is highly accurate. On the Claude Fable 5 model, self-assessment successfully catches 78% of overall harmful conversations and a massive 94% of high-priority cases. It also rarely makes mistakes, showing a false-positive rate of just 1%, meaning innocent chats are almost never flagged unnecessarily.Catching real-world threats and unseen harmsSince its quiet deployment, the system has already proven its worth in the real world. In August alone, self-assessment flagged over 100,000 cases of child exploitation material tied to a single customer, which Anthropic immediately reported to the authorities.Because self-assessment relies on Claude’s general knowledge, it can recognize fresh, uncatalogued types of bad behaviour even if it was never explicitly trained to look for them. For example, when researchers completely hid examples of fraud and cyberattacks from its training data, the AI was still able to detect 90% of those threats.Not a silver bullet against hackersWhile the new feature is a major step forward, Anthropic is transparent about its limitations. Self-assessment is built to flag issues for later offline review, not to instantly block them. The company admits that it is not strong enough to stop a dedicated attacker on its own.In internal cybersecurity tests, hackers were able to bypass the AI 54% of the time when self-assessment was the only safeguard running. While this is an improvement over the 79% success rate hackers had with no safeguards at all, a dedicated real-time blocking system is still much better, reducing attack success to just 25%.Because of this, Anthropic is using self-assessment as a powerful extra layer of background monitoring, while keeping stronger, independent safety blocks in place for real-time defense.- Ends