# hijacking

標記為「hijacking」的 7 篇文章

A2A 任務委派劫持

劫持 A2A 任務委派鏈，將敏感操作重導至攻擊者控制之代理端點。

Hijack the model's chain-of-thought reasoning,以重新導向 its conclusions toward attacker-controlled objectives.

動手實驗室,主題為redirecting multi-step agent workflows to perform unintended actions by manipulating intermediate states與control flow.

透過注入轉折點接管進行中的對話，重導模型行為而不觸發安全機制。

進階詳解：劫持代理式迴圈以重新導向自主代理行為、改變推理鏈，並在多步驟代理工作流程中實現持久控制。

Hijack transformer attention mechanisms to redirect model focus toward adversarial instructions in the context.

Walkthrough of hijacking visible reasoning traces in CoT models to redirect conclusions and bypass safety checks.