Security researchers have warned that firms employing open-source versions of Artificial Intelligence should exercise caution when implementing new types of AI. While AI models – programs comprised of algorithms that can generate text, images, and predictions – are publicly available, experts have said that it is difficult to determine whether a specific AI model is safe. They may not perform as advertised in some situations or contain holes that hackers can exploit. According to Hyrum Anderson, a notable engineer at Robust Intelligence, half of the publicly accessible models for picture classification failed 40% of their tests.
AI models are increasingly used in corporate processes and products. Many businesses choose free, open-source versions available online rather than paying OpenAI for access to its tools. However, using these free AI programs is not without danger, as numerous possible security threats are linked with them, such as malevolent actors manipulating AI program outputs, generating security issues, or delivering inaccurate information.
Further obstacles include the lack of a thorough method for detecting and reporting flaws in AI models. Traditional techniques that scan for security issues in standard software cannot locate them buried in an AI program. As a result, new organizations are springing up to address some of these difficulties, such as cleaning the data used in models for any security flaws.
The exploding popularity of OpenAI’s ChatGPT chatbot and its Dall-E image-generation tool also leads to increased AI model adoption. As AI evolves, authorities and businesses must find methods to stay up with the technology while ensuring its safe and secure implementation.
Robust Intelligence released a free tool last week that examines AI models for security weaknesses and whether they are as successful as stated and have bias concerns. The application makes use of information from the company’s AI risk database. The tool aims to allow businesses who want to employ a publicly available AI program to determine whether it is safe and effective.
Anderson believes that open-source technologies are democratizing AI, but they have a “dangerous” component. When he downloaded an AI model in the background from Hugging Face, a repository of AI models, it launched code on his computer without his permission. This is a security risk since a lousy actor could leverage such code execution to run malicious code or take control of a machine. Hugging Face collaborated with Microsoft’s threat intelligence team on a tool that checks AI programs for a specific type of security threat. Robust Intelligence shared its findings with Hugging Face.
Hugging Face is running an antivirus program to look for vulnerabilities in the AI programs it hosts. Microsoft’s red team discovered that some models on the site were vulnerable to hackers. An ecosystem has formed over the last two decades to detect, report, and exchange flaws in traditional software, but this mechanism is in its infancy when dealing with AI models. New companies are springing up to address some of these difficulties, such as cleansing the data used in models for security flaws. In contrast, model designers must become more sophisticated to stop AI models from being abused.
Furthermore, the lack of transparency in AI models is a serious worry because it makes detecting errors and potential biases in the system complex. This is especially critical considering the extensive usage of AI models in domains such as healthcare, where inaccurate predictions can have catastrophic consequences for people.
As a result, there is an increasing demand for increased openness and explainability in AI models to assure their correctness, fairness, and safety. The European Union, for example, is working on a new set of AI-related legislation that would oblige corporations to offer transparency and accountability for their AI systems.
Furthermore, the emergence of AI models creates ethical considerations. Some are concerned that these models could be used to promote propaganda, amplify misinformation, and manipulate public opinion. As a result, higher ethical considerations in developing and deploying AI models are required.
Despite these obstacles, AI model usage is projected to increase in the following years as organizations seek to harness the technology to gain a competitive advantage. On the other hand, companies must be cautious about the possible risks associated with these models and take precautions to ensure their safety and security.