Imagine a world where the software that powers your favorite apps, secures your online transactions, and keeps your digital life could be outsmarted and taken over by a cleverly disguised piece of code. This isn’t a plot from the latest cyber-thriller; it’s actually been a reality for years now. How this will change – in a positive or negative direction – as artificial intelligence (AI) takes on a larger role in software development is one of the big uncertainties related to this brave new world.

In an era where AI promises to revolutionize how we live and work, the conversation about its security implications cannot be sidelined. As we increasingly rely on AI for tasks ranging from mundane to mission-critical, the question is no longer just, “Can AI boost cybersecurity?” (sure!), but also “Can AI be hacked?” (yes!), “Can one use AI to hack?” (of course!), and “Will AI produce secure software?” (well…). This thought leadership article is about the latter. Cydrill (a secure coding training company) delves into the complex landscape of AI-produced vulnerabilities, with a special focus on GitHub Copilot, to underscore the imperative of secure coding practices in safeguarding our digital future.

You can test your secure coding skills with this short self-assessment.

The Security Paradox of AI

AI’s leap from academic curiosity to a cornerstone of modern innovation happened rather suddenly. Its applications span a breathtaking array of fields, offering solutions that were once the stuff of science fiction. However, this rapid advancement and adoption has outpaced the development of corresponding security measures, leaving both AI systems and systems created by AI vulnerable to a variety of sophisticated attacks. Déjà vu? The same things happened when software – as such – was taking over many fields of our lives…

At the heart of many AI systems is machine learning, a technology that relies on extensive datasets to “learn” and make decisions. Ironically, the strength of AI – its ability to process and generalize from vast amounts of data – is also its Achilles’ heel. The starting point of “whatever we find on the Internet” may not be the perfect training data; unfortunately, the wisdom of the masses may not be sufficient in this case. Moreover, hackers, armed with the right tools and knowledge, can manipulate this data to trick AI into making erroneous decisions or taking malicious actions.

Copilot in the Crosshairs

GitHub Copilot, powered by OpenAI’s Codex, stands as a testament to the potential of AI in coding. It has been designed to improve productivity by suggesting code snippets and even whole blocks of code. However, multiple studies have highlighted the dangers of fully relying on this technology. It has been demonstrated that a significant portion of code generated by Copilot can contain security flaws, including vulnerabilities to common attacks like SQL injection and buffer overflows.

The “Garbage In, Garbage Out” (GIGO) principle is particularly relevant here. AI models, including Copilot, are trained on existing data, and just like any other Large Language Model, the bulk of this training is unsupervised. If this training data is flawed (which is very possible given that it comes from open-source projects or large Q&A sites like Stack Overflow), the output, including code suggestions, may inherit and propagate these flaws. In the early days of Copilot, a study revealed that approximately 40% of code samples produced by Copilot when asked to complete code based on samples from the CWE Top 25 were vulnerable, underscoring the GIGO principle and the need for heightened security awareness. A larger-scale study in 2023 (Is GitHub’s Copilot as bad as humans at introducing vulnerabilities in code?) had somewhat better results, but still far from good: by removing the vulnerable line of code from real-world vulnerability examples and asking Copilot to complete it, it recreated the vulnerability about 1/3 of the time and fixed the vulnerability only about 1/4 of the time. In addition, it performed very poorly on vulnerabilities related to missing input validation, producing vulnerable code every time. This highlights that generative AI is poorly equipped to deal with malicious input if ‘silver bullet’-like solutions for dealing with a vulnerability (e.g. prepared statements) are not available.

The Road to Secure AI-powered Software Development

Addressing the security challenges posed by AI and tools like Copilot requires a multifaceted approach:

1. Understanding Vulnerabilities: It is essential to recognize that AI-generated code may be susceptible to the same types of attacks as „traditionally” developed software.
2. Elevating Secure Coding Practices: Developers must be trained in secure coding practices, taking into account the nuances of AI-generated code. This involves not just identifying potential vulnerabilities, but also understanding the mechanisms through which AI suggests certain code snippets, to anticipate and mitigate the risks effectively.
3. Adapting the SDLC: It’s not only technology. Processes should also take into account the subtle changes AI will bring in. When it comes to Copilot, code development is usually in focus. But requirements, design, maintenance, testing and operations can also benefit from Large Language Models.
4. Continuous Vigilance and Improvement: AI systems – just as the tools they power – are continually evolving. Keeping pace with this evolution means staying informed about the latest security research, understanding emerging vulnerabilities, and updating the existing security practices accordingly.

Navigating the integration of AI tools like GitHub Copilot into the software development process is risky and requires not only a shift in mindset but also the adoption of robust strategies and technical solutions to mitigate potential vulnerabilities. Here are some practical tips designed to help developers ensure that their use of Copilot and similar AI-driven tools enhances productivity without compromising security.

Implement strict input validation!

Practical Implementation: Defensive programming is always at the core of secure coding. When accepting code suggestions from Copilot, especially for functions handling user input, implement strict input validation measures. Define rules for user input, create an allowlist of allowable characters and data formats, and ensure that inputs are validated before processing. You can also ask Copilot to do this for you; sometimes it actually works well!

Manage dependencies securely!

Practical Implementation: Copilot may suggest adding dependencies to your project, and attackers may use this to implement supply chain attacks via “package hallucination”. Before incorporating any suggested libraries, manually verify their security status by checking for known vulnerabilities in databases like the National Vulnerability Database (NVD) or accomplish a software composition analysis (SCA) with tools like OWASP Dependency-Check or npm audit for Node.js projects. These tools can automatically track and manage dependencies’ security.

Conduct regular security assessments!

Practical Implementation: Regardless of the source of the code, be it AI-generated or hand-crafted, conduct regular code reviews and tests with security in focus. Combine approaches. Test statically (SAST) and dynamically (DAST), do Software Composition Analysis (SCA). Do manual testing and supplement it with automation. But remember to put people over tools: no tool or artificial intelligence can replace natural (human) intelligence.

Be gradual!

Practical Implementation: First, let Copilot write your comments or debug logs – it’s already pretty good in these. Any mistake in these won’t affect the security of your code anyway. Then, once you are familiar with how it works, you can gradually let it generate more and more code snippets for the actual functionality.

Always review what Copilot offers!

Practical Implementation: Never just blindly accept what Copilot suggests. Remember, you are the pilot, it’s “just” the Copilot! You and Copilot can be a very effective team together, but it’s still you who are in charge, so you must know what the expected code is and how the outcome should look like.

Experiment!

Practical Implementation: Try out different things and prompts (in chat mode). Try to ask Copilot to refine the code if you are not happy with what you got. Try to understand how Copilot “thinks” in certain situations and realize its strengths and weaknesses. Moreover, Copilot gets better with time – so experiment continuously!

Stay informed and educated!

Practical Implementation: Continuously educate yourself and your team on the latest security threats and best practices. Follow security blogs, attend webinars and workshops, and participate in forums dedicated to secure coding. Knowledge is a powerful tool in identifying and mitigating potential vulnerabilities in code, AI-generated or not.

Conclusion

The importance of secure coding practices has never been more important as we navigate the uncharted waters of AI-generated code. Tools like GitHub Copilot present significant opportunities for growth and improvement but also particular challenges when it comes to the security of your code. Only by understanding these risks can one successfully reconcile effectiveness with security and keep our infrastructure and data protected. In this journey, Cydrill remains committed to empowering developers with the knowledge and tools needed to build a more secure digital future.

Cydrill’s blended learning journey provides training in proactive and effective secure coding for developers from Fortune 500 companies all over the world. By combining instructor-led training, e-learning, hands-on labs, and gamification, Cydrill provides a novel and effective approach to learning how to code securely.

Check out Cydrill’s secure coding courses.