Anthropic's Claude silently blocks AI development tasks affecting 0.03% of traffic

Anthropic has quietly introduced silent limitations in its Claude model to prevent it from being effectively used for building competing AI systems. The interventions—ranging from prompt modification and steering vectors to parameter-efficient fine-tuning (PEFT)—operate in the background without notifying users. Unlike Anthropic's visible safeguards in cybersecurity, biology, and chemistry, these restrictions are invisible, meaning developers may unknowingly receive degraded outputs when working on frontier LLM development, pretraining pipelines, distributed training infrastructure, or ML accelerator design.

According to Anthropic, these changes affect only 0.03% of total traffic across fewer than 0.1% of organizations—a negligible fraction aimed squarely at actors already violating the Terms of Service by using Claude to build competing models. The rationale, outlined in Anthropic's February 2026 Risk Report, is that other AI developers are building powerful systems with similar risks but without the same safety standards. By silently throttling those most likely to ignore ToS, Anthropic aims to discourage unsafe development while maintaining a helpful user experience for the vast majority of legitimate applications.