CryptoAILab/Awesome-LM-SSP

Project name: Awesome-LM-SSP

What it is: Awesome-LM-SSP is a hand-curated, community-maintained collection (often called an "awesome list") of research resources about the trustworthiness of large AI models. It focuses on three big areas — Safety, Security, and Privacy (the "SSP" in the name) — and pays special attention to multi-modal models, like vision-language models (which process images and text together) and diffusion models (which generate images).

---

What's inside

The list organizes resources into several categories. As of the latest update, it includes:

• Papers (2,467 total) — scholarly articles grouped into three main branches:
  • Safety (1,207 papers) — subtopics like jailbreak attacks (tricks to bypass model restrictions), alignment (making models follow human intent), deepfakes, ethics, fairness, hallucination (when models invent facts), prompt injection (manipulating a model's instructions), and toxicity.
  • Security (521 papers) — subtopics like adversarial examples (inputs designed to fool a model), agent security, poison and backdoor attacks (planting hidden triggers during training), side-channel attacks, and system security.
  • Privacy (739 papers) — subtopics like membership inference attacks (figuring out whether a specific piece of data was used in training), model extraction (stealing a model's behavior), privacy-preserving computation, unlearning (removing data from a trained model), and watermarking / copyright.
• Surveys (42) — overview articles that summarize research in a field.
• Toolkits (33) — software packages for working with models.
• Leaderboards (5) — rankings of model performance.
• Competitions (5) — organized challenges.
• Books (3) — longer written references.

Each entry is tagged with badges that show, at a glance, what kind of model it involves (e.g., LLM, vision-language model, speech model, or diffusion model), what contribution it makes (e.g., new benchmark, new dataset, defensive technique), and where it was published (e.g., a conference, a blog, or a specific organization like OpenAI).

---

How it works

• The list is manually collected and explicitly labeled as "in progress" — the maintainers add items by hand.
• All entries are tracked in a shared Google Sheets database, which anyone can view. If a paper's status changes (say, an arXiv preprint gets accepted at a conference), people can update that information directly in the spreadsheet.
• To add a new paper, contributors fill out a submission table (with title, link, code, venue, classification, model type, and comments) via a GitHub issue.

---

Current status and community

• The project was first released on January 9, 2024, and celebrated its second anniversary on January 9, 2026 (the README includes birthday announcements).
• It is maintained by five named organizers and is inspired by several earlier similar projects (like LLM Security and Awesome LLM Security).
• The README includes a star-history chart, showing that the project welcomes community support and interest.

---

In plain terms: Think of it as a well-organized, frequently updated bibliography — with labels and filters — for anyone who wants to understand how large AI models can fail, be attacked, or leak private information, and what researchers are doing about it.