Anthropic has introduced new restrictions for its flagship Claude Opus 5.5 model, specifically targeting requests related to the development of frontier large language models. If the system detects queries concerning the creation of kernels for machine learning accelerators, it automatically redirects the conversation to the less capable Opus 5 model.
Similar safety mechanisms have been implemented for dual-use biological queries, including virology, toxicology, and molecular design. In these cases, Opus 5.5 utilizes classifiers similar to those found in the Fable model to trigger a fallback to Opus 5. Both models also block attempts to extract internal reasoning processes, such as requests to verbatim repeat chains of thought.
While the automatic model switching feature is enabled by default, users have the option to disable it in the settings, which results in the direct blocking of restricted queries. Anthropic states that these measures are intended to mitigate risks as artificial intelligence capabilities continue to expand.
For legitimate organizations in cybersecurity and life sciences, the company offers verification programs. Participants in the Cyber Verification Program and Life Sciences Verification Program gain access to more capable models with fewer restrictions. Currently, organizations using Opus 4.8 can access Opus 5 with reduced cybersecurity limitations.
