We Reproduced Anthropic's Mythos Findings with Public Models

Vidoc Security reproduced Anthropic's Mythos vulnerability research using public models (GPT-5.4 and Claude Opus 4.6), demonstrating that frontier AI vulnerability-discovery capabilities are spreading beyond internal labs. The researchers achieved exact reproductions on FreeBSD and Botan, exact on OpenBSD with Claude Opus 4.6 only, and partial results on FFmpeg and wolfSSL. The findings challenge the restriction-focused framing—the real advantage lies in validation, prioritization, and operationalization, not model access.