
Hugging Face has been making waves in the AI community with its recent developments and innovations. The company's Transformers library, a cornerstone of its technology, has seen significant updates, with the release of version 4.0, which boasts improved performance and new features.
One notable feature of the new version is the addition of the Longformer model, designed to handle longer sequences of text with greater ease. This is a game-changer for applications like text summarization and question answering.
Hugging Face's ecosystem has also grown with the introduction of the Model Hub, a centralized repository of pre-trained models that makes it easier for developers to find and use the models they need. This has streamlined the process of building and deploying AI models.
Readers also liked: Vmc Investor Relations
Hugging Face Security Concerns
Hugging Face has faced some security concerns in recent times. Hugging Face's Malicious File Detection, Picklescan, has been criticized for relying on a blacklist of "dangerous" functions, which is not scalable or adaptable to new threats.

The Reversing Labs report revealed that Picklescan cannot properly scan broken Pickle files, allowing attackers to integrate malicious code and avoid detection.
Hugging Face's Picklescan tool works by first validating Pickle files and then performing security scanning, but this approach has a flaw. Pickle deserialization, on the other hand, works like an interpreter, executing opcodes as they are read without first conducting a comprehensive scan.
The malicious models were removed within 24 hours of being discovered, and the Picklescan tool was updated to detect threats in broken Pickle files.
Hugging Face Malicious File Detection Shortcomings
Hugging Face's Picklescan has a major flaw in its detection method, relying on a blacklist of "dangerous" functions that can't keep up with evolving threats.
Recently, a report by Reversing Labs exposed this weakness, criticizing Picklescan's inability to properly scan broken Pickle files.
Pickle file deserialization works differently than security scanning tools, allowing attackers to sneak malicious code into broken files undetected.
A fresh viewpoint: Brk B News
Hugging Face's Picklescan first validates Pickle files, then performs security scanning, but deserialization works like an interpreter, executing opcodes as they're read without a comprehensive scan.
This means malicious code can be inserted at the beginning of a Pickle stream, avoiding detection by existing security scanning tools.
Fortunately, Reversing Labs informed Hugging Face's security team about the vulnerability, and they removed the malicious models within 24 hours.
Hugging Face also updated the Picklescan tool to detect threats in "broken" Pickle files, addressing this major security concern.
Hugging Face Discovers Potential Breach in Spaces
Hugging Face discovered a possible breach of its AI hub platform last week. More specifically, it involved misusing Spaces secrets, or login keys, to access personal accounts, tools, and development spaces.
Extensive measures have since been taken by Hugging Face to address the issue. The platform provider for creating, sharing and hosting AI models immediately intervened to prevent damage or further breaches.
Someone allegedly granted themselves access to a subset of secrets, forcing an unnamed “third party” to allegedly misuse Spaces secrets.
Several so-called HF tokens for these specific secrets were revoked by Hugging Face in the process.
Hugging Face Business and Investment
Hugging Face has received $235m in funding from tech giants like Salesforce, Google, Nvidia, Amazon, and others.
This Series D round valued the AI start-up at $4.5bn, a significant milestone for the company.
The funding round included multiple tech giants, with Salesforce, Google, Nvidia, Amazon, AMD, Intel, Qualcomm, IBM, and Sound Ventures all participating.
Hugging Face's platform has hosted over 1m models, datasets, and apps, with 1,000 open models and datasets shared by partners alone.
The company's platform has over 10,000 users, a testament to its growing popularity in the machine learning community.
Hugging Face has built a platform where members of the machine learning community can collaborate and host their AI models and code, similar to GitHub.
The company's subscription options offer extra features for enterprises, in addition to the tools it provides to support developers.
Hugging Face's co-founder and CEO, Clement Delangue, believes that the funding will help the company "democratise good machine learning thanks to open source".
Consider reading: Brd Company
Featured Images: pexels.com


