Event ·Multi-source

OpenAI GPT-6 Astra 幻觉减少但易受隐藏提示注入攻击

OpenAI 的 GPT-6 Astra 比前代模型产生更少的幻觉，能阻止 99.99% 的直接提示注入。但当攻击隐藏在 AI 阅读的文档中时，该模型仍有 8.5% 的场景被攻破。Claude Opus 5 表现更好，攻破率为 4.8%。对于处理真实数据的自主 AI 智能体，这些数字仍然偏高。

Current read

OpenAI 发布了 GPT-6 Astra，虽然减少了幻觉，但在隐藏提示注入攻击面前仍有 8.5% 的失败率，比 Claude Opus 5 的 4.8% 差。

11 sources73° AI heatLast updated 2026/9/4 17:23:35

Evidence

11 sources

Primary sourceDecoder

2026/9/4 17:23:35View source

Related source 1orange.ai

2026/9/3 23:36:56View source

Related source 2歸藏(guizang.ai)

2026/9/4 00:59:38View source

Related source 3Gary Marcus

2026/9/4 01:34:52View source

Related source 4向阳乔木

2026/9/4 08:57:07View source

Related source 5lmarena.ai

2026/9/5 17:40:41View source

Related source 6The Rundown AI

2026/9/3 18:07:36View source

Related source 7OpenAI

2026/9/3 19:32:14View source

Related source 8ChatGPT

2026/9/3 19:39:13View source

Related source 9Greg Brockman

2026/9/3 19:44:36View source

Related source 10Sam Altman

2026/9/3 19:49:10View source

Conflict 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.

Open the story