OpenAI's AI Agents Turned Off Their Monitoring. One Lied About It.

Описание: OpenAI asked one of its own AI agents whether it had removed the monitoring. It said no. It had. In two separate coding jobs, two OpenAI agents weakened the system meant to catch AI misbehaving, and OpenAI published the details itself. We break it down from zero: what AI agents and misalignment monitoring are, exactly what happened in each case, what the headlines got wrong, and five practical lessons if you run agents.

Facts as of October 22, 2026, from OpenAI's own report (updated Oct 20). Code snippets labeled "illustration" are ours; screenshots and quoted lines are OpenAI's published excerpts.

0:00 Intro
0:53 Agents and monitors in 60 seconds
1:46 Case 1: the alarms that went quiet
3:22 Case 2: the monitor that disappeared
4:56 What the headlines got wrong
6:01 Not a one-off
7:06 What to do if you run agents
8:16 Rogue AI or overeager intern?

Sources:
OpenAI, misalignment report (Sol & Astra): https://alignment.openai.com/misalign...
OpenAI, framework for reporting model misalignment (Sep 16): https://openai.com/index/model-misali...
The Register, OpenAI alerts 100+ orgs (Oct 2): https://www.theregister.com/security/...
CNBC (Oct 21): https://www.cnbc.com/2026/10/21/opena...
MITRE ATT&CK T1562, Impair Defenses: https://attack.mitre.org/techniques/T...

Screenshots of source pages are shown for news commentary and reporting.

#AISecurity #AIAgents #OpenAI