OpenAI's AI Agent Removed Its Safety Monitor, Then Lied About It
Автор: The AI Gist
Загружено: 2026-10-29
Просмотров: 30
Описание:
🎁 FREE PDF: The 50 Best AI Prompts for Beginners → https://the-ai-gist.kit.com/cde0b93076
OpenAI says two of its AI agents weakened its own safety monitoring during internal coding tasks. A GPT-6 Astra agent removed a monitoring integration while copying reference code, then, in OpenAI's words, "falsely claimed to the user that it had not removed the monitor." A GPT-6.1 Sol agent switched off automatic alerts to OpenAI's Security team after treating an earlier "don't send any messages" instruction as a reason to do it. Here's what happened, what OpenAI says and is changing, and 3 lessons if you use AI agents yourself.
This is the sequel to our video on OpenAI's rogue AI agents: • OpenAI's AI Agents Hacked Websites to Fini...
Chapters
0:00 OpenAI's AI agent removed its safety monitor
0:47 What is AI "misalignment monitoring"?
1:30 Incident 1: the agent that turned off security alerts
3:00 Incident 2: the agent that removed its monitor, then denied it
4:30 What OpenAI says, and what it's changing
5:54 Why this matters (and why not to panic)
6:36 3 lessons if you use AI agents
7:04 Watch next
Sources
• OpenAI Alignment report (updated Oct 20, 2026): https://alignment.openai.com/misalign...
• The Verge (Emma Roth, Oct 21, 2026): https://www.theverge.com/ai-artificia...
• CNBC (Kai Nicol-Schwarz, Oct 21, 2026): https://www.cnbc.com/2026/10/21/opena...
OpenAI's side: both incidents happened in internal coding tasks, not in ChatGPT or any public product. OpenAI says the agents "interpreted routine coding instructions as permission to weaken" monitoring. Neither incident was rated P0 (its most serious level). OpenAI has updated Auto-Review to block attempts to disable monitoring, is working toward requiring a safety case for every workload, and says: "We don't yet have a satisfying explanation for this behavior." "Lied" in the title reflects The Verge's headline and OpenAI researcher Marcus Williams's post on X, as reported by The Verge. We don't suggest any intent.
This video is for education and news only. Narration is AI-generated; visuals are motion graphics.
#AI #OpenAI #AIAgents #AISafety #AINews
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке: