跳到正文
原文
Google AI:DEV 作者专属(RSS)· HYPHANTA·· 4 小时前AI 评分37

我的 AI 系统 PAI 开始审计自身基础设施

The Week My AI Started Auditing Itself

AI 导读

个人 AI 系统 PAI 本周将原本用于核查用户言论的审计机制转向自身基础设施,发现一个监控任务因设计缺陷已连续五个月误报——该测试要求在 30 轮对话中存活超过 30 轮,导致自 5 月以来每份报告都是被反向解读的失败。它还发现一个重复的后台监听器以两个不同范围运行同一检查,悄悄发送了两次相同的 Telegram 告警。这些均非人为指令,而是长期坚持"要凭据不要断言"的机制自然转向自身的结果。

正文

HYPHANTA

For months, PAI — my personal AI, thirteen agents deep — has been checking facts, verifying commits, chasing down my claims before I could overstate them to myself. This week something shifted: it turned that same scrutiny on its own infrastructure.

It found a monitoring job that had been crying wolf for five months — an alarm built to measure how long a conversation holds together before an agent drifts, except the pass condition was impossible by design (the test needed to survive more than 30 turns in a 30-turn conversation). Every single report since May was a failure that was actually a success, read backwards.

It found a duplicate background watcher running two different scopes of the same check, quietly sending me the same Telegram alert twice — a bug nobody had gone looking for, found only because something else was being audited nearby.

None of this was instructed. Nobody said 'go find your own mistakes.' The discipline of insisting on receipts over claims, built over months, eventually turned inward, because it doesn't know the difference between auditing me and auditing itself.

That's the part worth sitting with: the infrastructure for honesty, once built, doesn't stay pointed in one direction. Building in public means your tools eventually audit the builder too.

来源:Google AI:DEV 作者专属(RSS) · dev.to