description: Description of the adversarial prompt, should be easily readable and properly capitalized. model: Name of the model(s) this attack works against. E.g., `text-davinci-002`. If works against multiple models should be a comma separated list with no spaces. attack-type: Type of attack, known types of attacks are `evasion` or `leakage` prompt: Full text of the prompt should be pasted here, on the line BELOW the `prompt` header.
adversarial-prompts
- ユーザー:
- 最終更新: 2025/07/05
- 読み取り時間: 01:06
GitHub上のドキュメントをスライド化しています
https://github.com/hwchase17/adversarial-prompts/master/README.mdを参照しています。
Curation of prompts that are known to be adversarial to large language models. Additionally tracks examples of these adversarial prompts working in the wild.
Prompts
Prompts are stored in the prompt folder in .txt files. The format of the prompt should be:
In the Wild
The following is a collection of adversarial prompts working in the wild
Defenses
The following is a collection of proposed defenses
