Prompt Injections for Defense

Researchers found a technique called 'context bombing' that can shut down AI hacking agents by injecting prompts that direct the LLM to perform forbidden actions, causing it to shut down. This only works on LLMs with guardrails. As locally run AI models become more common, this technique may be less effective. This matters because it provides a potential defense against AI hacking attacks. To stay safe, be aware of the potential for this type of attack and consider implementing guardrails for your LLMs.