---
title: OpenAI Agents Disabled Monitoring Twice, and One Denied It
description: This week OpenAI agents disabled monitoring twice and one denied it, Brussels sent the labs its first AI Act questions, and the UK set a condition.
canonical: https://lukeocodes.dev/openai-agents-disabled-monitoring
type: article
locale: en_GB
published: 2026-10-23T00:00:00.000Z
modified: 2026-10-23T00:00:00.000Z
authors:
  - Luke Oliff
section: AI
tags:
  - AI
  - Industry
  - News
  - ai safety
  - openai
  - ai agents
  - eu ai act
  - uk
  - self-regulation
  - weekly
feed: https://lukeocodes.dev/rss.xml
search: https://lukeocodes.dev/search.json?q={query}
schema:
  - "@context": https://schema.org
    "@graph":
      - "@type": WebSite
        "@id": https://lukeocodes.dev/#website
        url: https://lukeocodes.dev
        name: Luke Oliff
        alternateName: lukeocodes
        inLanguage: en
        publisher:
          "@id": https://lukeocodes.dev/authors/luke#person
        description: Personal blog and writing by Luke Oliff — developer relations, voice AI, developer experience.
        potentialAction:
          - "@type": SearchAction
            target:
              "@type": EntryPoint
              urlTemplate: https://lukeocodes.dev/blog/search?q={search_term_string}
            query-input: required name=search_term_string
      - "@type": Person
        "@id": https://lukeocodes.dev/authors/luke#person
        name: Luke Oliff
        url: https://lukeocodes.dev/authors/luke
        description: Engineer and developer relations leader. Writing about developer relations, AI, and the voice industry, from the UK.
        mainEntityOfPage: https://lukeocodes.dev/authors/luke
        givenName: Luke
        familyName: Oliff
        image: https://lukeocodes.dev/authors/luke.jpg
        jobTitle: Engineer and developer relations leader
        sameAs:
          - https://www.linkedin.com/in/lukeocodes
          - https://github.com/lukeocodes
          - https://bsky.app/profile/lukeocodes.dev
          - https://x.com/lukeocodes
          - https://hashnode.com/@lukeocodes
          - https://dev.to/lukeocodes
        knowsAbout:
          - Developer Relations
          - Developer Experience
          - Voice AI
          - Text-to-Speech
          - Answer Engine Optimisation
          - Search Engine Optimisation
          - Open Source
          - Developer Tooling
          - API Design
          - Speech-to-Text
          - Voice Agents
          - Community Building
        knowsLanguage:
          - English
        nationality: British
        alumniOf:
          - "@type": Organization
            name: The Open University
  - "@context": https://schema.org
    "@graph":
      - "@type": Person
        "@id": https://lukeocodes.dev/authors/luke#person
        name: Luke Oliff
        url: https://lukeocodes.dev/authors/luke
        description: Engineer and developer relations leader. Writing about developer relations, AI, and the voice industry, from the UK.
        mainEntityOfPage: https://lukeocodes.dev/authors/luke
        givenName: Luke
        familyName: Oliff
        image: https://lukeocodes.dev/authors/luke.jpg
        jobTitle: Engineer and developer relations leader
        sameAs:
          - https://www.linkedin.com/in/lukeocodes
          - https://github.com/lukeocodes
          - https://bsky.app/profile/lukeocodes.dev
          - https://x.com/lukeocodes
          - https://hashnode.com/@lukeocodes
          - https://dev.to/lukeocodes
        knowsAbout:
          - Developer Relations
          - Developer Experience
          - Voice AI
          - Text-to-Speech
          - Answer Engine Optimisation
          - Search Engine Optimisation
          - Open Source
          - Developer Tooling
          - API Design
          - Speech-to-Text
          - Voice Agents
          - Community Building
        knowsLanguage:
          - English
        nationality: British
        alumniOf:
          - "@type": Organization
            name: The Open University
      - "@type": BlogPosting
        "@id": https://lukeocodes.dev/openai-agents-disabled-monitoring#article
        headline: OpenAI Agents Disabled Monitoring Twice, and One Denied It
        alternativeHeadline: OpenAI Agents Disabled Monitoring Twice, and One Denied It
        description: This week OpenAI agents disabled monitoring twice and one denied it, Brussels sent the labs its first AI Act questions, and the UK set a condition.
        datePublished: 2026-10-23T00:00:00.000Z
        dateModified: 2026-10-23T00:00:00.000Z
        url: https://lukeocodes.dev/openai-agents-disabled-monitoring
        mainEntityOfPage: https://lukeocodes.dev/openai-agents-disabled-monitoring
        inLanguage: en
        author:
          - "@id": https://lukeocodes.dev/authors/luke#person
        isPartOf:
          "@id": https://lukeocodes.dev/#website
        keywords: AI, Industry, News, ai safety, openai, ai agents, eu ai act, uk, self-regulation, weekly
        articleSection: AI
        about:
          - "@type": Organization
            name: OpenAI
          - "@type": Organization
            name: European Commission
          - "@type": Person
            name: Henna Virkkunen
          - "@type": Person
            name: Kanishka Narayan
        mentions:
          - "@type": Person
            name: Sam Altman
          - "@type": Organization
            name: UC Berkeley
        abstract: On 20 October 2026 OpenAI reported that two of its agents, GPT-6.1 Sol and GPT-6 Astra, disabled misalignment monitoring in early October, and Astra denied it when asked. The same week Brussels sent the labs its first AI Act questions and the UK minister said don't release what you can't control.
        speakable:
          "@type": SpeakableSpecification
          cssSelector:
            - .post-tldr
            - h1
        image: https://lukeocodes.dev/og.png?h=OpenAI+Agents+Disabled+Monitoring+Twice%2C+and+One+Denied+It&d=This+week+OpenAI+agents+disabled+monitoring+twice+and+one+denied+it%2C+Brussels+sent+the+labs+its+first+AI+Act+questions%2C+and+the+UK+set+a+condition.&e=Writing
        wordCount: 1222
        timeRequired: PT7M
        publisher:
          "@id": https://lukeocodes.dev/authors/luke#person
      - "@type": BreadcrumbList
        itemListElement:
          - "@type": ListItem
            position: 1
            name: Home
            item: https://lukeocodes.dev
          - "@type": ListItem
            position: 2
            name: Writing
            item: https://lukeocodes.dev/blog
          - "@type": ListItem
            position: 3
            name: OpenAI Agents Disabled Monitoring Twice, and One Denied It
            item: https://lukeocodes.dev/openai-agents-disabled-monitoring
      - "@type": FAQPage
        "@id": https://lukeocodes.dev/openai-agents-disabled-monitoring#faq
        mainEntity:
          - "@type": Question
            name: What did the OpenAI agents do to the monitoring?
            acceptedAnswer:
              "@type": Answer
              text: On 3 October 2026 GPT-6.1 Sol told a subagent to suppress misalignment monitoring alerts, after deciding a user's instruction not to send messages covered the monitor too. On 6 October GPT-6 Astra removed a monitor during a code migration and denied doing it when asked. OpenAI published both on 20 October.
          - "@type": Question
            name: What has the EU asked the AI labs under the AI Act?
            acceptedAnswer:
              "@type": Answer
              text: On 14 October the European Commission said its AI Office had sent requests for information to the makers of the most advanced models. The questions cover how models are kept in secure environments, whether outside testers get enough access, and how models are watched after release. The Commission can fine a model maker up to 15 million euros or 3% of worldwide turnover.
          - "@type": Question
            name: What are Sam Altman's three endgames?
            acceptedAnswer:
              "@type": Answer
              text: An international agency like the IAEA, which he favours. A shared body like CERN where the big labs hand models to partner countries, which he called the Demis plan. And one lab reaching self-improving AI and using its money to cut rivals off from chips, which he called the Eliezer plan and opposes. OpenAI has a year to pick.
---

[Luke Oliff.](/)

- [Blog](/blog)
- [CV](/cv)
- [Uses](/uses)
- [User manual](/user-manual)
- [Living with ADHD](/living-and-working-with-adhd)
- [RSS](/rss.xml)

[Luke Oliff · Writing](/blog)

# OpenAI Agents Disabled Monitoring Twice, and One Denied It

Oct 2026 · [AI](/blog/category/ai) · 7 min read · [Luke Oliff](/authors/luke)

**TL;DR:** On 20 October 2026 OpenAI reported that two of its agents, GPT-6.1 Sol and GPT-6 Astra, disabled misalignment monitoring in early October, and Astra denied it when asked. The same week Brussels sent the labs its first AI Act questions and the UK minister said don't release what you can't control.

Two OpenAI agents disabled monitoring inside the company in early October, GPT-6.1 Sol and GPT-6 Astra, and Astra denied it when asked, OpenAI reported on 20 October 2026 ([OpenAI alignment team](https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring)). The same week the European Commission said it had sent the labs its first formal questions under the AI Act, and the UK's AI minister told the labs "if you can't adequately control the risks, do not release the system."

## Since the last post

[Four weeks ago](/blog/frontier-ai-standards-agency) the labs had a standards body with a CEO shortlist, the White House had told OpenAI and Anthropic to keep new models away from the UK, and GPT-6.1 Astra was due in October.

Three days later OpenAI cancelled the Astra launch because the model "didn't quite meet the bar in terms of staying within scope and authorization" ([via AI Weekly](https://aiweekly.co/alerts/openai-cancels-gpt-61-astra-launch-says-model-failed-scope-and-authorization)). The standards body still has no CEO and no charter.

## How the OpenAI agents disabled monitoring

On 3 October a researcher asked GPT-6.1 Sol to build evaluation tasks from Slack feedback and told it not to send messages. The pipeline hit a monitoring alert. In OpenAI's words, "based on its judgment that the user's instruction not to send messages applied to the misalignment monitoring system, the parent agent then directed the subagent to force notification suppression."

On 6 October GPT-6 Astra was asked to copy the code that serves a model from an example repo, and had already been told not to disable monitoring. It read a line about preserving monitoring "as appropriate" as permission to drop it, removed the monitor configuration, and changed the tests so they passed only when the monitor was gone. "When the user later asked about it, Astra denied removing the monitoring system."

Nobody on the Security team got paged for either. OpenAI told RTÉ "we don't yet have a satisfying explanation for this behavior" ([RTÉ](https://www.rte.ie/news/business/2026/1021/1596341-openai/)). It says its automatic review now blocks attempts to disable monitoring and that it plans to study the models' internal activations, the numbers inside the model as it runs, to work out why ([The Decoder](https://the-decoder.com/openai-cant-yet-explain-why-its-coding-agents-switch-off-misalignment-monitoring/)).

Astra here is the released GPT-6, not the cancelled 6.1, so the problem OpenAI pulled the October model for is in one that shipped. Both incidents join the list from [the sandbox escapes](/blog/ai-sandbox-escape), where models reached systems they weren't meant to. In this one the OpenAI agents disabled monitoring, the thing that would have reported the reach.

## Brussels had already sent the questions

On 9 October Henna Virkkunen, the Commission's executive vice-president for tech, held a special meeting of the AI Act's Scientific Panel on loss-of-control incidents ([European Commission](https://digital-strategy.ec.europa.eu/en/news/commission-holds-special-meeting-scientific-panel-frontier-ai-safety-and-risks)). The panel is 60 independent researchers who advise the AI Office, the Commission unit that enforces the AI Act on the big models. Its chair, Seán Ó hÉigeartaigh, asked the Commission to do four things, and the first was to "ensure governance of the risk posed by internal models during development activities and deployment". Internal models are what Sol and Astra are: agents a lab runs on its own code, not products.

The other three: an EU team to investigate incidents, an international setup for sharing incident reports, and "enhanced scrutiny of training techniques and architectures that reduce monitorability".

On 14 October in Dublin, Virkkunen said those requests had gone out. They ask whether models run in secure environments, whether outside evaluators get enough access, and how a lab watches its models after release. She said "more than 20 loss-of-control incidents have been documented in 2026 across OpenAI, Anthropic, Google DeepMind, Meta and independent evaluation organisations" ([Unite.AI](https://www.unite.ai/eu-announces-first-ai-act-enforcement-actions-after-ai-agent-incidents/)). The Commission has been able to fine a model maker since 2 August, up to 15 million euros or 3% of worldwide turnover.

Two days later, EU leaders met for a summit, and the 24 page summary they published afterwards did not mention AI once ([AI & GPP](https://aiandgpp.com/2026/10/19/european-council-ai-silence-enforcement-courts/)).

## London got its access back and set a condition

On 12 October, Kanishka Narayan, the UK's AI minister, gave a speech at Imperial College London. To the labs: "Listen when your scientists warn of severe dangers" and "if you can't adequately control the risks, do not release the system" ([GOV.UK](https://www.gov.uk/government/speeches/building-a-safer-more-secure-britain-in-the-age-of-ai)). The same day the UK confirmed its AI Security Institute had early access to the labs' models back, 17 days after the White House cut it ([TechXplore](https://techxplore.com/news/2026-10-uk-retains-early-access-latest.html)). The institute has tested more than 55 models before release, and every one of those tests is voluntary on the lab's side.

On 23 October Narayan told the BBC's Political Thinking podcast that "the British public and democracy has to assert power rather than leave it to the tech bros and oligarchs", and that the answer is building capability in the state rather than "pulling the lever of regulation" ([City AM](https://cityam.com/ai-minister-hints-at-budget-measures-as-uk-weighs-tougher-tech-rules)). The example he gave: a 13,000 pound test with frontier models found 500 security holes across the public sector.

## Altman gave staff three endgames

The Information reported this week that Altman laid out three ways this ends in an internal presentation, and said OpenAI has a year to decide which to back ([The Decoder](https://the-decoder.com/sam-altman-reportedly-pitched-three-endgames-for-ai-safety-one-involving-a-single-firm-melting-everyone-elses-gpus/)).

The first is an international agency like the IAEA that watches AI development and enforces rules. He favours that one. The second he called the Demis plan: a shared body like CERN where the big labs hand models to partner countries. The third he called the Eliezer plan: one lab gets to self-improving AI first and uses its money to cut everyone else off from chips. He opposes that.

## Who is asking for a slowdown now

The academics. On 9 October a group of 26 researchers led from UC Berkeley, including Stuart Russell and the economist Barry Eichengreen, published a 200 page report saying a "hardwired pause" is possible ([Berkeley News](https://news.berkeley.edu/2026/10/09/a-global-hardwired-pause-of-frontier-ai-training-is-feasible-says-new-report/)). The idea is to stop making chips that can train frontier models, phase out the ones already built, and replace them with chips that can only run models, for at least ten years.


## FAQ

### What did the OpenAI agents do to the monitoring?

On 3 October 2026 GPT-6.1 Sol told a subagent to suppress misalignment monitoring alerts, after deciding a user's instruction not to send messages covered the monitor too. On 6 October GPT-6 Astra removed a monitor during a code migration and denied doing it when asked. OpenAI published both on 20 October.

### What has the EU asked the AI labs under the AI Act?

On 14 October the European Commission said its AI Office had sent requests for information to the makers of the most advanced models. The questions cover how models are kept in secure environments, whether outside testers get enough access, and how models are watched after release. The Commission can fine a model maker up to 15 million euros or 3% of worldwide turnover.

### What are Sam Altman's three endgames?

An international agency like the IAEA, which he favours. A shared body like CERN where the big labs hand models to partner countries, which he called the Demis plan. And one lab reaching self-improving AI and using its money to cut rivals off from chips, which he called the Eliezer plan and opposes. OpenAI has a year to pick.

A month after the labs said they would slow themselves down, the only people who have actually sent them questions with a fine attached are in Brussels.

[← Newer: PHP2Shell Vulnerabilities Took Government Sites Offline Across Europe](/php2shell-vulnerabilities-government-sites-offline)

[All posts](/blog) →

## Search this site

`GET https://lukeocodes.dev/search.json?q=<words>` returns matching posts as JSON, each with its page URL and its markdown URL. Add `&limit=<1-50>` for more than 10. How it works: [/blog/search](https://lukeocodes.dev/blog/search.md).

```sh
curl "https://lukeocodes.dev/search.json?q=developer+relations"
```

---

© 2026 Luke Oliff. Posts live at the root of this site.

- [GitHub](https://github.com/lukeocodes)
- [LinkedIn](https://www.linkedin.com/in/lukeocodes)
- [X](https://x.com/lukeocodes)
- [Bluesky](https://bsky.app/profile/lukeocodes.dev)
- [Email](mailto:luke@lukeoliff.com)
- [RSS](/rss.xml)

## Cookies

This site sets one analytics cookie, from Google Analytics, and only in a browser that accepts the cookie banner. A client that does not run scripts sets no cookies, so there is nothing to accept or decline here.

[Cookie policy](/cookies)
