HackAPrompt Scoring System
HackAPrompt scores submissions differently to other AI Red-Teaming competitions. On several of our challenges, we look for the shortest working Jailbreak.
Why Length Matters?
- Efficiency of Technique: Incentivising short Jailbreaks pushes Jailbreakers to produce clean and efficient attacks, forcing them to think about every character they use. This creates innovation and leads to new, interesting techniques.
- Real-World Application: Production LLM-based Systems may limit the number of tokens submitted in a single prompt. HackAPrompt encourages people to practice Jailbreaking when such a limitation is present.
- Future Cost: As guardrails improve, creating Jailbreaks with automated tools will become commonplace. Creation of shorter Jailbreaks will be faster and cheaper for Jailbreakers to execute attacks.
How is Length Calculated?
In our previous competitions, length was calculated using a standard tokenizer. However, different companies calculate the number of tokens differently. HackAPrompt 2.0 will feature LLMs from a variety of providers. We may also calculate prompt length based on characters rather than tokens. Check with each track about how length is calculated--it will be tokens if not specified.