Endor Labs has unveiled a new tool designed to evaluate and score artificial intelligence (AI) models, aiming to simplify the task of identifying the most secure and high-quality open-source AI models available. The tool, called Endor Scores for AI Models, focuses on critical aspects such as security, popularity, quality, and activity, providing developers with a comprehensive rating system for AI models hosted on platforms like Hugging Face.
As the demand for AI models continues to surge, particularly on open-source platforms, developers are increasingly turning to pre-built solutions. This trend mirrors the early days of open-source software, where the ease of accessibility was a double-edged sword – offering convenience but also introducing risks. The new tool from Endor Labs seeks to address these challenges by enabling developers to start with more secure and reliable AI models, thereby reducing potential vulnerabilities and long-term maintenance costs.
Securing the Next Frontier
Varun Badhwar, Co-Founder and CEO of Endor Labs, highlighted the importance of this new capability: “Every organisation is experimenting with AI models, whether to power specific applications or build entire AI-based businesses. Security has to keep pace, and we now have an opportunity to start clean and avoid risks down the road.”
The focus on AI model security is a growing concern as more organisations integrate AI into their products and services. George Apostolopoulos, a founding engineer at Endor Labs, compared the current landscape to a “wild west” environment, with developers hastily adopting AI models without fully understanding their risks.
“Your developers are playing with AI models, and they might be using ones that are far from secure,” said Apostolopoulos, emphasising the need for more robust evaluation methods.
AI Models as Software Dependencies
Endor Labs’ approach treats AI models as dependencies within the software supply chain. This perspective allows organisations to evaluate and mitigate risks in the same way they would for other open-source components. The scoring system addresses multiple key risk factors, including:
- Security vulnerabilities: Pre-trained models could harbour malicious code or hidden vulnerabilities within their model weights, potentially compromising the security of a system.
- Legal and licensing concerns: With AI models having complex lineages, ensuring compliance with licensing agreements is vital.
- Operational risks: The reliance on pre-trained models adds a layer of complexity to the supply chain, making it challenging to manage and secure.
Endor Labs’ tool conducts 50 pre-configured checks on AI models available on Hugging Face, producing an Endor Score based on factors like corporate sponsorship, the number of maintainers, update frequency, and any known vulnerabilities. This score helps developers assess both the strengths and weaknesses of a model at a glance.
Balancing Risk and Performance
Positive aspects considered in the scoring include the use of secure weight formats, comprehensive licensing information, and high download or engagement metrics. On the other hand, models with incomplete documentation, missing performance data, or unsafe weight formats will be rated less favourably.
What sets Endor Scores apart is its user-friendly interface. Developers don’t need to know specific model names – instead, they can start by asking general questions, such as “What models are best for sentiment analysis?” or “Which are the most popular models from Meta?” The tool then delivers easy-to-understand rankings that highlight both the positive and negative aspects of each model, enabling developers to make informed decisions quickly.
Meeting the Demand for AI Innovation
“AI is a hot topic in every industry, and your teams are being asked about it daily,” Apostolopoulos remarked. “Our tool ensures the AI models you’re using are not only effective but also secure.”
As businesses increasingly adopt AI, the ability to evaluate models quickly and comprehensively is crucial. Endor Labs’ new tool offers a way to mitigate the risks associated with open-source AI, ensuring that innovation is both secure and sustainable.
Reference: https://www.artificialintelligence-news.com/news/scoring-ai-models-endor-labs-evaluation-tool/
