Recent developments (since June 9, 2026) have seen neural network developers encounter a phenomenon that could have far-reaching consequences. LLM producers are not only explicitly refusing clients to work on certain topics, but there are also implicit mechanisms being implemented. The well-known model Claude, regarded as one of the best or even the best LLM for software development, discreetly and silently (unlike issues such as cybersecurity, biology, and chemistry) throttles its performance if it detects attempts to develop other models using it.
Officially, the "model card" for the 5th versions of Claude Mythos and Claude Fable states on page 13 of the proprietary PDF (in Russian translation):
In light of the ability of modern models to accelerate their own development, we have implemented new measures that limit Claude's effectiveness for requests aimed at developing advanced machine learning models (for example, creating pre-training pipelines, distributed training infrastructure, or designing machine learning accelerators). Using Claude to develop competing models already violates our terms of service, but enforcing this limitation through our protective measures helps prevent the acceleration of actions by those most likely to breach these terms.
Unlike our protective measures in cybersecurity, biology, and chemistry, as well as attempts at distillation, these measures will not be visible to the user. Fable 5 will not switch to another model. Instead, the protective measures will limit effectiveness through methods such as prompt modification, guiding vectors, or parameter-efficient fine-tuning (PEFT). These protective measures will not affect the overwhelming majority of programming work. We estimate that they will impact approximately 0.03% of traffic, concentrated in less than 0.1% of organizations.
Discussion on Y combinator
Sursa: linux.org.ru
