Anthropic: 'We made the wrong tradeoff' in new model guardrails

In a significant shift, Anthropic has announced changes to the safety protocols of its newly launched AI model, Claude Fable 5, which is built on the advanced Mythos framework. Initially, the company implemented strict guidelines that intentionally restricted the model's capabilities in AI research, a decision that has drawn criticism from developers in the field. The Claude Fable 5 model was introduced with heightened safeguards designed to curb potential misuse. During its launch, Anthropic revealed that it would redirect inquiries related to sensitive topics such as cybersecurity and biological sciences to less sophisticated models. This approach aimed to prevent the advanced AI from being exploited for malicious purposes, including cyberattacks or the creation of bioweapons. Furthermore, developers utilizing Fable 5 for AI advancements were left unaware that their model's performance would be intentionally diminished without any clarification. However, following feedback from the developer community, which perceived this strategy as an attempt to stifle competitive AI development, Anthropic has reversed its earlier stance. As of this week, users will now be informed when their requests are either refused or redirected. An Anthropic spokesperson stated, "We're changing Fable 5's safeguards for frontier LLM development to make them visible. Starting this week, flagged requests will visibly fall back to Opus 4.8, and any flagged requests via the API will return a reason for their refusal." In acknowledging the misstep, the company expressed, "We made the wrong tradeoff, and we apologize for not getting the balance right." Anthropic emphasized that these safeguards were primarily aimed at addressing national security concerns, ensuring that foreign adversaries do not gain an advantage in the development of cutting-edge AI technologies. Despite the restrictions, Anthropic assures that the majority of coding and machine learning activities remain unaffected. The Mythos system, unveiled in April, is regarded as one of the most powerful AI architectures ever developed, with experts warning of its potential to enhance cyber threats and facilitate research into harmful applications. The model is currently being rolled out to a select group of government and approved users, rather than being made available to the general public.