Primary sourceIT之家
View sourceEvent ·Multi-source
Anthropic自曝安全漏洞:生物武器过滤器失效近一年,涉及1.33亿次对话
Anthropic在8月14日发布安全报告,承认其拦截化学、生物武器相关危险请求的分类器在2025年5月至2026年4月期间失效。约5万名外部承包商在此期间产生1.33亿次与模型的对话,未经过相关生物安全分类器过滤。内部调查未发现这些请求被实际用于滥用的证据,但公司已加强承包商筛查要求。此前Anthropic在2025年5月发布Claude Opus 4时启用了ASL-3防护措施。
Current read
Anthropic自己承认生物武器过滤器失效快一年,5万承包商产生1.3亿次对话没被拦。虽然没发现滥用,但安全防线有洞,值得看看。
11 sources73° AI heatLast updated 2026/8/16 11:37:52
Evidence
11 sourcesRelated source 1Gary Marcus
View sourceRelated source 2宝玉
View sourceRelated source 3Decoder
View sourceRelated source 4AI Will
View sourceRelated source 5Anthropic
View sourceRelated source 6techcrunch
View sourceRelated source 7Dario Amodei
View sourceRelated source 8Claude Code: GitHub Releases
View sourceRelated source 9lmarena.ai
View sourceRelated source 10爱范儿
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.