Anthropic confirms Claude Opus 5 embeds invisible safeguards — prompt modification, steering vectors, PEFT — specifically to limit its usefulness for training f
Anthropic has confirmed that its Claude Opus 5 model incorporates advanced, invisible safeguards designed to prevent its misuse for training other large language models. These technical measures, including prompt modification and steering vectors, operate beneath the user-facing prompt layer. This approach raises questions about the auditability and external verification of these safety features. AI
IMPACT These advanced, invisible safeguards could set a new standard for model safety, potentially influencing how other labs approach AI security and auditability.