红队测试Red Teaming
Red Teaming
发布前雇人主动攻击模型找漏洞：安全声明背后的实际工序。Hiring people to attack a model before release to find holes: the real process behind safety claims.
红队测试指以对抗性方式测试系统、找出潜在漏洞。
Anthropic 官方表述：红队是提升 AI 安全的关键工具。
前沿实验室在部署前做内部评测，并引入外部红队（含众包平台）。
Anthropic 自己也承认：红队缺乏标准化实践，各家做法差异大。
Red teaming means testing a system adversarially to identify potential vulnerabilities.
Anthropic's official framing: red teaming is a critical tool for improving AI safety.
Frontier labs run internal evaluations before deployment and bring in external red teamers, including crowdworker platforms.
Anthropic itself concedes red teaming lacks standardized practice; approaches differ widely.
「经过安全测试」的含金量取决于红队怎么做的。知道没有统一标准，你才会去问测试细节。'Tested for safety' is only as strong as how the red teaming was done. Knowing there is no uniform standard makes you ask for details.
去导航：找现成的工具、平台、评测与学习资源Directory: ready-made tools, platforms, benchmarks and courses ↗
回到名词地图看第2站的其他 35 个词 →Back to the map: the other 35 terms in station 2 →
- Anthropic 官方《Challenges in red teaming AI systems》Anthropic: Challenges in red teaming AI systems ↗