Using LLMs to secure source code

使用大语言模型保障源代码安全

Anthropic describes a six-step Claude Opus workflow for source-code security: threat modeling, sandboxing, vulnerability discovery, validation, triage, and remediation; in its open-source scanning work, it disclosed 1,596 vulnerabilities by May 22, 2026, with 97 already fixed.

Why it matters: HKR-H/K/R all pass: Anthropic gives a Claude Opus security-audit workflow plus 1,596/97 outcome numbers. It stays below 85 because this is not a new model or platform-level capability release.