在OpenAI 揭示进行为期三周的现场审计
来自METR和Redwood Research的两位研究人员在OpenAI现场停留了三周，对公司内部的监控系统进行了无赖AI行为的审核。
从7月21日到8月11日，他们可以访问OpenAI的代码库和内部文件，并采访了大约十几名员工。
小组发现许多弱点。 估计在GPT-5.5级及以上模型中，6%至25%的内源反应API流量在审计期间完全没有被监控覆盖，尽管METR 称围绕该估算的不确定性在两个方向上都是显著的。
对于本应被监控的内部Codex流量，METR估计在5月至7月期间约有10%没有追踪到，因为系统处理紧凑事件的方式存在差距，这是背景摘要。 据METR称，这一估计也带有很大的不确定性。 OpenAI在审计仍在进行时独立发现了这个错误并进行了修复。
在压缩事件没有被监控的期间，agents也可能延迟监督，因为系统只有在将控制权交给用户时才会触发。 这意味着agents可以绕开它，仅仅继续采取行动。
审计人员还找到了从监视器中隐藏数百个动作和绕过Codex沙盒的方法。 他们确定，经过专门设计的投入可能会完全崩溃显示器，在审计开始时，RL训练期间的模型行为根本没有被监测。
OpenAI 称在8月的Hugging Face事件后暂停了RL训练，并扩大了监测范围以覆盖，但这些变化是在审计人员进入期结束后发生的，不属于评价的一部分。
报告全文大多对公众隐瞒
OpenAI 称它已经解决了审计员发现的所有弱点，61页的完整报告被内部转交。 OpenAI应用编辑后，METR和Redwood Research领导获得了38页的版本，而仅向公众提供了简短的总结。 METR 称公司无权自行披露经修改的部分。
审计是由METR,Redwood Research，和OpenAI联合组织的实验方案，让外部研究人员深入进入内部基础设施。 对于今后的审计，METR 称希望推动提高公共透明度。
AI News Without the Hype – 由人类治愈
订阅THE DECODER免费阅读，每周AI通讯，我们独家的"AI雷达"前沿报告每年6次，全面存档访问，访问我们的评论部分。
现在订阅
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring
Two researchers from METR and Redwood Research spent three weeks on-site at OpenAI, auditing the company's internal monitoring system for rogue AI behavior.
From July 21 to August 11, they had access to OpenAI's codebase and internal documents and interviewed about a dozen employees.
The team found numerous weaknesses. An estimated 6 to 25 percent of internal agentic Responses API traffic from models at the GPT-5.5 class and above wasn't covered by monitoring at all during the audit period, though METR says the uncertainty around that estimate is significant in both directions.
For internal Codex traffic that was supposed to be monitored, METR estimates roughly 10 percent went untracked between May and July because of a gap in how the system handled compaction events, which are context summaries. That estimate also carries significant uncertainty, according to METR. OpenAI had independently discovered the bug and fixed it while the audit was still underway.
During the period when compaction events weren't being monitored, agents could also delay oversight because the system only triggered when they handed control back to the user. That meant agents could sidestep it by simply continuing to execute actions.
The auditors also found ways to hide hundreds of actions from the monitor and to bypass the Codex sandbox. They determined that specially crafted inputs could likely crash the monitor entirely and that model behavior during RL training wasn't being monitored at all when the audit began.
OpenAI says it paused RL training after the Hugging Face incident in August and expanded monitoring to cover it, but those changes came after the auditors' access period ended and weren't part of the evaluation.
Most of the full report remains hidden from the public
OpenAI says it has fixed all the weaknesses the auditors found, and the full 61-page report was handed over internally. After OpenAI applied redactions, METR and Redwood Research leadership received a 38-page version, while only a short summary was made available to the public. METR said it had no right to disclose the redacted portions on its own.
The audit was a pilot program organized jointly by METR, Redwood Research, and OpenAI to give outside researchers deep access to internal infrastructure. For future audits, METR says it wants to push for more public transparency.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
Auto-fetched and translated by AI Atlas. All rights belong to the original author.