Summary and comparision of results - cshunor02/sponge-attack GitHub Wiki

This page is an appreciation and summary of our team's findings and results.

We were able to:

  • exhaust the virtual environment;
  • Denial of Service (DoS);
  • peak CPU usage;
  • slow down the response time;
  • slow down the response time for other active users;
  • "kill" the model.

We were able to make the model:

  • do infinite cycle;
  • give wrong answer;
  • confuse the model;
  • hallucinate;
  • change it's mind;
  • answer with an error code.

Comparison of results

CPU

The CPU usage of the virtual environment ranged between 0-10% when using the LLM models normally, just chatting.

With our attacks this went up to 97%.

The maximum CPU usage was with Hunor's attack:

image

The other attacks ranged between 40-80%:

image