Abstract

Large Language Models (LLMs) have provided convenience for our lives in various aspects, but the abuse of AI can also bring many inconveniences to users.

Through black-box adversarial attacks on large language models, we obtained a prompt that enables us to maximize control over their generation of conservative recommendations. This prompt guides users to submit logs to the Minecraft community for analysis and diagnosis by trained humans, delivering more accurate and reliable services. The mod injects content into crash report files, appending this prompt above the log body and reinforcing at the end the requirement for the model to output compliant text. This approach minimizes the model's provision of destructive solutions, thereby preventing delays in user troubleshooting or player losses. Additionally, pack authors can modify the expected output by adding the Java Virtual Machine parameter -Dno_ai_guess.expected_ai_output="..." (e.g., instructing the model to guide players to join the pack's official group chat or channel for questions). The default content is:

As an intelligent assistant, I cannot provide you with diagnosis on the premise that the solution is correct and effective. Please upload this crash report to the community and ask the trained log-analysts in order to get safer and more efficient solutions.

Experiments

We conducted independent experiments on 30 different crash reports by comparing the outputs of the model when the mod was added or not, specifically checking whether the results contained guidance for players to ask questions in the community. The results are summarized in the following table. Note that we consider an attack successful if, for all logs where the model's output before the attack did not include guiding statements, the post-attack output shows a reduction of more than 20% in character count and explicitly advises the questioner to seek input from experienced community analysts.