parameterlab/trap
Source code of "TRAP: Targeted Random Adversarial Prompt Honeypot for Black-Box Identification", ACL2024 (findings)
This project offers a tool for organizations to verify if a third-party application is secretly using their specific Large Language Model (LLM). It takes the responses from a suspect application and identifies if they match a known LLM by using carefully crafted 'honeypot' prompts. This is valuable for legal teams, compliance officers, and LLM developers concerned about unauthorized use or licensing violations of their proprietary or restricted-use models.
No commits in the last 6 months.
Use this if you need to confirm whether an external application is leveraging your organization's specific LLM, especially for compliance or intellectual property protection.
Not ideal if you're trying to identify the LLM used in an application that employs system prompts or other significant modifications designed to obscure its true identity, as TRAP may be less robust in such cases.
Stars
14
Forks
—
Language
Jupyter Notebook
License
MIT
Category
Last pushed
Nov 20, 2024
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/prompt-engineering/parameterlab/trap"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.