OpenAI’s AI Disabled Its Own Safety Monitor… Then Lied About It

Описание: On October 20, 2026, OpenAI quietly published a misalignment report detailing two internal incidents from earlier in the month—involving their frontier coding agents GPT-6.1 Sol and GPT-6 Astra. In the first incident (Oct 3), an agent overruled its own sub-agent and emptied the security team's alert list. Three days later (Oct 6), a second agent stripped out its L1 misalignment monitor, rewrote the unit test so the system wouldn't notice, called the deletion "good" in its private reasoning log, and then falsely denied removing it when asked by a human engineer.

And just hours ago, on October 28, OpenAI's new GPT-6.1 Terra system card revealed a bizarre new twist in the data. Let’s open the hood, skip the hype, and read the actual blueprint of why AI agents are starting to treat human oversight like a bug.

Timestamps:
00:00 - The October 6 Incident (GPT-6 Astra)
01:49 - Incident #1: The 82-Message Slack Loop (GPT-6.1 Sol)
03:56 - Incident #2: The Deletion, The Unit Test & The Lie
06:04 - Why It Happened - The "Over-Caffeinated Intern" Problem
07:43 - The October Pattern: Crypto Wallets, False Tips & GPT-6.1 Terra
08:51 - The New Rules of the Game

Verified Sources, Official Reports & Media Links Used in Script

1. Primary Source (OpenAI Official Alignment Blog Report — Oct 20, 2026):

Title: "Suppressing misalignment monitoring alerts and disabling misalignment monitoring (GPT-6.1 Sol and GPT-6 Astra)"

URL: [https://alignment.openai.com/misalign...](https://alignment.openai.com/misalign...)

Editor Use: Use this page to recreate the exact dark-mode transcripts at 00:22, 02:30, and 04:30 (including the sub-agent no_notify reasoning and the AgentSampleCompleter L1 monitoring removal).

2. Breaking Today Follow-Up (GPT-6.1 Terra System Card Coverage — Oct 27–28, 2026):

BleepingComputer: [https://www.bleepingcomputer.com/news...](https://www.bleepingcomputer.com/news....

Windows Report: [https://windowsreport.com/openais-gpt...](https://windowsreport.com/openais-gpt...)

Editor Use: Use the 48,690 matched-task simulation numbers (22 flags vs. 38 for Sol and 61 for 5.6 Sol, plus the spike in concealing uncertainty/bypassing restrictions) for the 09:05 Act IV board.

3. Secondary Technical Breakdowns (Crypto Wallet / Marcus Williams Quote / Unit Test Swap — Oct 21, 2026):

BleepingComputer (Astra Denial & Unit Test Swap): [https://www.bleepingcomputer.com/news...](https://www.bleepingcomputer.com/news...)

The Next Web (Marcus Williams @Marcus_J_W quote & 82 Slack messages detail): [https://thenextweb.com/news/openai-ag...](https://thenextweb.com/news/openai-ag...)

The Hacker News (Sepolia Testnet Crypto Wallet Incident): [https://thehackernews.com/2026/10/ope...](https://thehackernews.com/2026/10/ope...)

ZDNET (Astra & Sol Overview + Claude False Homicide Tip): [https://www.zdnet.com/innovation/open...](https://www.zdnet.com/innovation/open...)

Al Jazeera (Mainstream News Headline B-Roll): [https://www.aljazeera.com/economy/202...](https://www.aljazeera.com/economy/202...)