AI That Can Hack? Anthropic Tested Mythos — Here’s What It Found

In a world driven by artificial intelligence, the fears of AI have already started to surface. While AI can take away jobs, it might also be powerful enough to hack software on its own. Anthropic has recently introduced its latest AI model, Mythos, which can detect critical vulnerabilities in software before hackers exploit them. However, Mythos has been making headlines for entirely different reasons. AI researcher at Anthropic, Nicholas Carlini, tested the company’s new AI model and realised the system could do far more than he expected, and not all of it was safe. Can Mythos Behave Like A Hacker?According to Bloomberg, Carlini found during his testing that the Mythos could break into digital infrastructure easily. He realised that the system could go

far beyond just assisting humans by acting on its own. Carlini and his team even found that Anthropic’s AI could identify system issues and exploit them on its own. It could even build its own hacking tools and target software like Linux. This simply means that instead of helping a hacker, the tool could itself behave like one. How Anthropic’s Leadership Reacted?Logan Graham, Anthropic’s Red Team leader, stated that his team immediately sensed something unusual and highlighted how quickly the model could show risky capabilities. Anthropic’s chief science officer and co-founder, Jared Kaplan, tricked the model since its origin and has been monitoring its progress. He explained to the team to decide whether these issues were minor or something that could actually affect the border public. His conclusion was that the AI could be risky. Kaplan and the co-founder Sam McCandlish presented their concerns to CEO Dario Amodei and President Daniela Amodei. Kaplan stated, “It quickly became clear that this wasn’t going to be a normal launch.”

The company has taken a safer path to release the model. It limited the access of Mythos to a select group of organisations under a programme called ‘Project Glasswing’. Anthropic’s idea is to use these AI tools as a defensive mechanism and help companies to spot and fix vulnerabilities before hackers find them.