Update
AI-Assisted Bug Hunting Increases Disclosures by 120%
A preprint study found that firms given early access to Anthropic's AI model Claude Mythos Preview experienced a 120% relative increase in software-flaw disclosures compared to Chinese firms without access. No detectable slowdown in repair channels was observed. The study covers publicly disclosed vulnerabilities and specific repair channels, but does not measure undisclosed workloads or user installation of patches.