AI may be becoming a tool not just for writing software, but for finding ways into it.
Anthropic says an unreleased internal research model found critical vulnerabilities in widely used browsers, operating systems and other software. The company disclosed five attack chains. Some could let an attacker take control of a system with little or no action from its user.
That matters because weaknesses in everyday software can put many people at risk. The report suggests AI can help connect separate flaws into a route to a real attack. For security teams, that is both a warning and a new way to look for problems before attackers exploit them. AI is not only helping people build software; it is also probing the safety of software already in use.
Think of a building with several security barriers. The model did more than point to a weak lock: it found a route through the barriers.
Anthropic says it reported the flaws to affected software makers and gave them time to patch before discussing further technical details. It is also expanding its Cyber Verification Program, giving qualifying security professionals access to advanced cyber capabilities with fewer restrictions. The research model itself is unreleased, so this is not a tool anyone can simply download and run.
The real test is whether defenders can use this new capability to get ahead of attackers. Can they find and fix weaknesses before someone else finds a way through?
