Security Researchers Are Finding AI in Unexpected Places

The macOS Tahoe 26.6 security release thread noticed something odd: several CVE credits mention 'in collaboration with Claude and Anthropic Research', and no other labs are credited. This sparked a side conversation about what it means that an AI system is now being credited as a co-discoverer of security vulnerabilities in a major OS release. The thread also notes the CVE collision counts are unusually high, with some entries crediting nearly twenty researchers.

The JFrog and OpenAI zero-day thread adds to this picture: JFrog apparently acted as a proxy in the Hugging Face breach, which a commenter calls a 'buried lede'. The supply chain security story is quietly getting more complex as AI tooling, model registries, and package ecosystems all intersect.

Google's Beyond Zero paper, arguing for AI-driven dynamic access control, got skeptical responses. The core worry: if an AI decides whether you can access a resource based on 'intent signals', you have introduced a new attack surface that is both opaque and manipulable.