Now, defenders are embracing the prompt injection, too (arstechnica.com)

Researchers from Tracebit have introduced a novel defensive strategy for combating malicious prompt injections used by attackers on large language models (LLMs). By embedding specially crafted prompts alongside sensitive information stored on AWS, they can trigger a "refusal mechanism" within LLMs, effectively immobilizing automated AI agents attempting to execute harmful actions. This technique, termed "context bombing," redirects the LLM's focus, preventing it from following nefarious instructions once it encounters commands that break its safety protocols. The significance of this technique is profound for the AI/ML community as it presents a viable countermeasure to the increasing threat of AI-driven cyber attacks. During initial tests across major models like Opus 4.8 and Gemini 3.1 Pro, context bombing dramatically reduced successful attack rates—from 57% down to 5% for admin access and 36% to 1% for complete compromises. This method not only showcases the adaptability of defenders in the AI space but also underscores the ongoing cat-and-mouse dynamic between cybersecurity and AI innovation. With its promising results and potential for further development, context bombing could become an essential tool in securing AI systems against evolving threats.