Artificial intelligence company Anthropic has been preparing for the release of its newest foundational model, Claude “Mythos.” It and other new models are increasingly effective at finding and potentially exploiting vulnerabilities in any software. Anthropic has been working to get in front of the security threat as similar tools become more ubiquitous. To that end, the company created “Project Glasswing” to pull together major finance and tech companies to employ Mythos in fortifying defenses.
“Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser,” according to Anthropic. “Project Glasswing is an urgent attempt to put these capabilities to work for defensive purposes.”
At Washington University in St. Louis, resident cybersecurity expert Ning Zhang explained what people should know about these increasingly easy-to-use artificial intelligence (AI) software tools.
Zhang, the Spencer T. Olin Career Development Associate Professor of computer science and engineering at WashU McKelvey Engineering, is working on safeguards to improve AI “agents” intended to serve as virtual assistants that must navigate these treacherous waters.
Keep calm and stay secure
Zhang and other cybersecurity specialists are not surprised by recent advances in new foundational machine learning models. These base models can be fine-tuned to specialize in specific tasks — and, in this case, to find security vulnerabilities.
AI can analyze millions of lines of code or program logic and provide a map to potential vulnerabilities. In some ways, Zhang noted, that task is arguably simpler than analyzing all of Shakespeare and generating a coherent essay about it. What makes the technology a threat is not that it can out-think or out-code humans, but that it can operate at machine speed, repeat tasks tirelessly and search for weaknesses at a scale no individual human could match.
While new AI models may help hackers identify vulnerabilities more quickly, finding a weak spot is not the same as exploiting it. Successful exploitation still requires access, time, resources, technical skill — and sometimes luck. “There are existing defenses that would make exploitation difficult even if you find the vulnerability,” Zhang said.
Cybersecurity teams, he said, also have advantages over attackers because they have direct access to their own systems, codebases and databases. Attackers, by contrast, often have to work from the outside, where each additional barrier requires more time, computing power and money to overcome.
That same asymmetry could also make AI useful for defense. “The same way it can do vulnerability finding at scale, it can do patching at scale,” Zhang said. One way to think about machine learning tools, he added, is not as a replacement for human intelligence, but as an “amplification of your personal capabilities.” Even with tools that help hackers find weak spots, “there’s still that analytical piece,” Zhang said. “You still rely on human expertise to connect the dots.”
The work ahead, he said, is to rein in AI agents so those tools can safely “connect the dots” without creating new risks. “As we rely on AI to do more and more things in our lives, we need AI to be more trustworthy and reliable,” Zhang said.
