A persistent security flaw known as 'Disregard that!' attacks threatens the core functionality of large language models (LLMs). This vulnerability allows attackers to inject malicious instructions directly into an LLM's context window, bypassing any built-in guardrails. The problem stems from the fundamental nature of the context window - the entire input text the model processes before generating output.

For chatbots, this includes the entire conversation history; for coding assistants, it encompasses code, style guides, and documentation; for search-like tools, it involves queries and retrieved documents. Disregard that! attacks exploit this openness, tricking models into ignoring prior instructions and executing harmful commands. Management's proposed solution - adding more defensive instructions to the context window - is fundamentally flawed, creating an unsustainable arms race. The vulnerability isn't limited to chatbots; any LLM processing untrusted material from APIs, search results, or file shares is at risk.

Even multi-agent systems fail to provide a solution, as compromised context windows can propagate malicious instructions between agents. The shutdown of OpenAI's Sora video generation app, while not explicitly stated, highlights the real-world consequences of this unresolved issue.