Claude Opus 5.5 keeps switching to Opus 4.8 (www.xda-developers.com)

Anthropic's Claude Opus 5.5, launched on September 22, has introduced a significant security feature where requests related to cybersecurity automatically switch responses to the older Opus 4.8 model. This safeguard prevents the advanced capabilities of Opus 5.5 from being exploited for harmful purposes. However, tests have shown that the fallback model, Opus 4.8, has been more susceptible to security breaches, as nearly 12.01% of rerouted requests were compromised compared to a mere 0.07% for the flagship model. This raises concerns about the reliability of the fallback method, as it could be exploited by attackers specifically looking to initiate harmful requests. The automatic handoff is triggered not only by specific prompts but also by the context of the materials the model processes, including memory and web searches. Consequently, users are encouraged to manage the switch carefully; options to disable the auto-switching feature allow for better control. Anthropic’s Cyber Verification Program further extends a verification pathway for users engaged in security work, emphasizing a tiered access system for advanced modeling capabilities. This development highlights the ongoing challenges in maintaining cybersecurity within AI systems and underscores the importance of rigorous testing and user controls to mitigate potential risks.