SALLIE: Safeguarding Against Latent Language & Image Exploits

SALLIE is currently a fresh research project (11 days old) with 0 stars and 3 forks, likely serving as the code accompaniment to the cited arXiv paper. While its goal of a 'modal-agnostic' defense is technically sophisticated and addresses a major pain point (unified safety for VLMs), it faces extreme competition. Frontier labs (OpenAI, Anthropic, Google) view safety and prompt injection mitigation as core platform requirements; they are building these capabilities natively into their model alignment (RLHF) and system prompts. Furthermore, established AI security startups like Lakera, HiddenLayer, and Robust Intelligence, along with NVIDIA's NeMo Guardrails, already provide more mature, production-ready alternatives. The defensibility is low because the project lacks a community, data moat, or unique infrastructure layer; it is an algorithmic approach that can be easily absorbed or superseded by a minor update to a frontier lab's safety filter. The high platform domination risk reflects the reality that security is increasingly becoming a 'feature' of the model provider rather than a standalone third-party tool.