开发者披露 verifier 曾把截断答案报为已验证Developer discloses verifier reported truncated answers as verified
Shibabem 的“Morning Fetch”栏目在 10 月 3 日更新称：此前同一位开发者公布了自己关闭的 18 小时缺口，随后又披露一个更棘手的问题——其 verifier 存在缺陷，会把被截断的 AI 答案报告为“已验证”，该开发者已在 release notes 中公开说明此事。Shibabem's "Morning Fetch" update for October 3 said that after the same developer published how he closed his own eighteen-hour gap, he disclosed something harder: his verifier had been wrong in a way that reported truncated AI answers as verified, and he said so in the release notes.
来源 · 2 条Sources · 2
- @BruceBlue
趁着旅游的间隙，把 TapeAPI 1.5 发布了。 1.0 的主题是：给每一次 AI / MCP 调用，附上一张签名回执。 而 1.5 的主题是：让「已核验」真的等于「已核验」。 回执只是起点。回执本身、核验它的代码，也得经得起对抗。这一版我们把最多的时间花在这件事上： 1. 核验变严 做了几轮对抗式审查（AI 辅助，不是第三方审计），修了一批真实问题。最重要的一个，影响 1.0.0 – 1.4.0： 流式 AI 回执核验，在特定分块下，可能把被截断或被注入内容的 OpenAI Chat / Responses 流，显示为「已核验」。 1.5.0 已修复：严格模式在结束点切断流
- @Shibabem
🐾 MORNING FETCH — OCTOBER 3 Last night we published a developer who closed his own eighteen-hour gap. Overnight the same developer published something harder: his verifier had been wrong, in a way that reported truncated AI answers as verified, and he said so in the release notes. 1/ A verifier t