Anthropic taught Cloud to identify dangerous requests independently Anthropic introduced a self-assessment mechanism that allows Cloud to independently identify signs of potentially dangerous use of artificial intelligence. After each response, the model analyzes the correspondence and assesses its security. In Cloud Devil 5, the system detects 78% of potentially dangerous dialogues and 94% of high priority cases with a false positive rate of about 1%. Since its introduction in June, the mechanism has helped identify serious problems, including more than 100,000 cases of child exploitation per customer, as well as attempted romance scams. Suspicious dialogues are not automatically blocked, but sent for further verification by Bloomberg.

Anthropic taught Cloud to identify dangerous requests independently Anthropic introduced a self-assessment mechanism that allows Cloud to independently identify signs of potentially dangerous use of artificial intelligence. After each response, the model analyzes the correspondence and assesses its security. In Cloud Devil 5, the system detects 78% of potentially dangerous dialogues and 94% of high priority cases with a false positive rate of about 1%. Since its introduction in June, the mechanism has helped identify serious problems, including more than 100,000 cases of child exploitation per customer, as well as attempted romance scams. Suspicious dialogues are not automatically blocked, but sent for further verification by Bloomberg.