Claude claims zero-day findings in Ghost and Linux kernel during 90-minute demo

• Anthropic researcher Nicholas Carlini reportedly showed a scaffolded Claude setup finding a zero-day in Ghost during a live demo, shifting the discussion from abstract capability claims to a named target and a bounded workflow demo thread.
• In the account summarized in the talk link post, the system found a blind SQL injection in about 90 minutes, extracted an admin API key, and then "repeated the same move" against the Linux kernel.
• The same thread says Anthropic's setup used a "surprisingly minimal scaffold" and had already uncovered "500+ high-severity vulnerabilities," though the public evidence here is a conference talk summary rather than a vendor writeup or disclosure note demo thread.

According to the primary thread, Carlini's demo centered on two concrete case studies: Ghost CMS and a Linux kernel bug in NFS. The Ghost example is the sharper engineering claim. The summary says Claude found a blind SQL injection in Ghost, a project described there as having about 50,000 GitHub stars and no prior critical vulnerability history, then used that path to take an admin API key.

The same account, echoed in the linked-talk post, frames the result as a capability threshold: a model with a "minimal scaffold" autonomously discovering and exploiting bugs in heavily audited software within a 90-minute session. That matters because the claim is no longer just that frontier models score better on security evals; it is that a scaffolded agent reportedly completed an end-to-end vulnerability workflow against disclosed software targets. What remains public, though, is still second-order evidence around a conference presentation and the talk video, not a detailed technical paper or reproduction package.