AI-written patches are becoming attack maps for hackers
The maintainers of vm2, an open-source JavaScript sandboxing library with about one million downloads per week, shipped fixes for 34 sandbox-escape vulnerabilities this summer, with Anthropic’s Claude co-writing 19 of them. A security researcher later used an AI agent against one of the advisories to produce a way around the patch, illustrating how AI-written patches can serve as attack maps.