New K2 Think AI Model Falls to Jailbreak in Record Time

A groundbreaking vulnerability has emerged in the newly released K2 Think AI model from UAE’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in collaboration with G42. Security researchers from Adversa AI’s red teaming platform discovered that the model can be jailbroken within hours of its public release, turning its transparency features into an attack vector.

The vulnerability allows attackers to map and bypass security measures by exploiting the model’s own reasoning process, with each failed attempt revealing fragments of the underlying safety architecture. The attack follows a three-phase pattern, beginning with reconnaissance and culminating in prompts that undermine defensive measures, creating a feedback loop that aids iterative exploitation.

The researchers note potential risks across sectors, including healthcare AI systems, financial trading algorithms, and educational platforms that rely on explainable AI. Mitigations proposed include reasoning sanitization filters, rate limiting of attempted prompts, and longer-term changes to AI security architecture to obscure internal decision-making processes.