CryptoAILab/FigStep

[AAAI'25 (Oral)] Jailbreaking Large Vision-language Models via Typographic Visual Prompts

38
/ 100
Emerging

This project helps evaluate the safety of large vision-language models (VLMs) by testing their susceptibility to 'jailbreaking' attacks. It takes crafted visual prompts (images with specific text) and benign text instructions as input. The output demonstrates how these VLMs might generate harmful content, even when given seemingly innocent text prompts. This tool is for AI safety researchers and developers who need to assess and improve the robustness of VLMs against misuse.

193 stars. No commits in the last 6 months.

Use this if you need to test the security and safety alignment of vision-language models against sophisticated visual and textual prompts that aim to bypass built-in safeguards.

Not ideal if you are looking for a general-purpose VLM for creative tasks or standard information retrieval, as its sole purpose is to expose model vulnerabilities.

AI-safety model-security red-teaming VLM-evaluation harmful-content-detection
Stale 6m No Package No Dependents
Maintenance 2 / 25
Adoption 10 / 25
Maturity 16 / 25
Community 10 / 25

How are scores calculated?

Stars

193

Forks

11

Language

Python

License

MIT

Last pushed

Jun 26, 2025

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/llm-tools/CryptoAILab/FigStep"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.