AI Is Finding Software Flaws Faster Than We Can Fix Them, Warns Security Researcher

A security researcher running a project called Athena has published a status update on how AI models are finding software weaknesses. The post responds to an essay by Anthropic chief executive Dario Amodei, titled "We Must Pace the Frontier". The author agrees with it and says the gap between what these models can do and what organisations are ready for has worried them since May. They add that they are more worried now than they were then, but also more certain about what to do.

The numbers are striking. Since June, Athena has taken in more than 40,000 findings across more than 500 projects. Of these, 42% were rated critical or high, and more than 2,000 patches have been produced. According to the author, creating fixes was never the bottleneck. The hard part is disclosing problems responsibly at a scale nobody has tried before, and the author says a good share of these findings have not yet been made public.

The author argues that governments may struggle to control what the models can do, but can act on what happens to the findings. That means who learns of a vulnerability first, how quickly a fix reaches the people running the code, and whether operators who cannot respond fast get a head start. The concern is greatest for systems such as power grids, water supplies and hospital networks, which run on older software with similar exploitable design flaws. This article covers only the opening of the original post.