Pliny the Liberator claims universal jailbreak of models (twitter.com)

Pliny the Liberator has announced the discovery of a universal jailbreak technique that reportedly allows for the circumvention of protections across all AI models, including flagship systems like Opus 5 and GPT-5.6 Sol. This technique is claimed to be highly effective and resistant to complete patching, which raises significant concerns regarding AI safety and security. While the developer has opted not to open-source this method immediately, citing the need for a responsible disclosure period, they are encouraging dialogue with industry experts in AI safety, policy, and alignment. The implications of this discovery are profound for the AI/ML community, as it challenges the integrity of model safety protocols and presents potential risks for misuse. By withholding public access to the jailbreak, Pliny aims to mitigate the likelihood of overcorrection in regulatory responses, advocating for discussions among stakeholders to explore its consequences and enhance understanding. This situation underscores the ongoing tension between innovation and safety in AI development, highlighting the critical need for responsible practices in managing emerging technologies.