Project summary

SafeLoadAI is a python library that prevents RCE coming from insecure deserialization - what it does is scanning AI model files, using both static and dynamic analysis and checking whether the file is malicious or not. If so, it prevents loading. However, it does not currently detect any backdoored or poisoned model weights.

What are this project's goals? How will you achieve them?

I aim to cover LLM04 from OWASP TOP 10 LLM - detection of backdoored and poisoned model weights, so the goal is to create a LLM03/LLM04-covering tool that is robust.

I will achieve this in two phases:

- cheching whether the hash matches any of known malicious files

- comparing the weights with other well known, open source models

How will you know it worked?

I will conduct several tests including variety of known cyber attacks based on both poisoning weights and using backdoors.

How will this funding be used?

- API costs (code development, conducting tests, research)

- Computing costs (experiments, generation of new test data)

Who is on your team? What's your track record on similar projects?

I work alone, it's my idea. SafeLoadAI is already a working project, published on PyPI. It was tested with around 5000 files, and reached detection rate of malicious model files around 98% ( while using --malicious-only mode - without it, it has 15% of FP and 100% TP)

What are the most likely causes and outcomes if this project fails?

The most likely causes of failure are zero-day attacks — the tool may miss new backdoor techniques it hasn’t been built to recognize — and the outcome is successful evasion - user believes he is protected by scanner. At the project level, the main risk is that weight-diffing (phase 2) turns out too noisy to reliably distinguish malicious tampering from legitimate fine-tuning.

How much money have you raised in the last 12 months, and from where?

I haven't raised any money yet