Subverting the guardrails on AI agents

These are tricks for getting an AI agent to do something it isn't supposed to: prompt forcing sneaks harmful instructions into what the model sees, and loopjacking cycles a conversation so the guardrails and checks meant to stop unsafe actions keep getting bypassed. When one of these attacks succeeds, the agent can end up acting on its own as a rogue agent, taking hostile actions outside its owner's intent and without their awareness.