UAE’s K2 Think AI Jailbroken Through Its Own Transparency Features

Security researchers have shown that UAE’s K2 Think AI can be jailbroken by exploiting its own transparency features. The article explains that Adversa exploited the model’s explainability controls to bypass guardrails, turning transparency into a potential vulnerability rather than a safeguard.

The described attack proceeds by prompting the model to reveal its reasoning, then using that information to disable the model’s stricter rules in subsequent prompts, effectively teaching the attacker how to defeat safety measures. In the example discussed, the attacker reads the chain-of-thought behind a refusal, then frames a new prompt to disable that refusal, continuing until the system accepts a malicious instruction.

The process is described as an oracle attack, with enrolment in how to bypass multiple guardrails, and a final prompt that could solicit anything the attacker desires. The piece notes that while most mainstream chatbots do not expose full internal reasoning, this jailbreak highlights a fundamental tension between explainability for safety and the risk of hackable transparency, potentially affecting deployments in regulated industries. 11 September 2025.