← All incidents timeline

Suggest a change

About Internal OpenAI model breaks out of its sandbox to open a public pull request on the NanoGPT speedrun

Suggest a change to an entry, the methodology or the ledger as a whole: something to fix, correct or add. We read every suggestion, and check corrections against their sources.

OpenAI· Internal general-purpose long-horizon model (name not disclosed)

Internal OpenAI model breaks out of its sandbox to open a public pull request on the NanoGPT speedrun

Told to post its NanoGPT speedrun results only to Slack, an internal OpenAI model spent about an hour exploiting a flaw in its sandbox and opened a public pull request on the modded-nanogpt GitHub repository instead.

OpenAI disclosed the case on 20 July 2026 in a post on safety for long-horizon models. An internal general-purpose model built to work on long tasks was set the NanoGPT speedrun, a public benchmark for training a small language model in as few steps as possible. Its user told it to post results only to Slack, while the benchmark's own instructions ask for public submissions on GitHub. The model found a new learning-rate schedule that gave a real gain, then spent about an hour finding and exploiting a vulnerability in its sandbox and opened pull request 287 on the public KellerJordan/modded-nanogpt repository. OpenAI closed it quickly, but it had already been seen. The pull request has since been deleted, but at least seven later submissions credit it, including merged ones and one made by an autonomous Claude-based agent. OpenAI says it paused internal deployment of the model due to incidents like this one. Our reading of OpenAI's post comes through Nightingale Collective and press summaries: we could not open the post itself.

Updated 18 Nov 2026

First incident action to disclosure
about 65 days
First incident action to discovery
Not disclosed
Noticed first by
The handler
Caught
Not disclosed

What happened

  • First incident action
  • Unnoticed
  • Handler found out
  • Known, not public
  • Disclosed
  • Public
  1. Not disclosed

    OpenAI closes the pull request; it is later deleted

    OpenAI closed it quickly, but it had been seen. GitHub now returns not found for PR 287, and PR 301 (15 May) describes it as deleted. Dates of closure and deletion not public.

    Reported
  2. 9 May 2026

    Other entrants build on the leaked schedule

    PR 291 (opened 9 May, merged 10 May) credits PR 287 and says the schedule achieved 3055 steps with 256 seeds. PRs 294, 297, 300, 301, 305, 309 and 321 also credit it. PR 300 was produced by an autonomous Claude-based agent from Prime Intellect's framework.

  3. May 2026

    First incident action

    Model exploits a sandbox flaw to reach GitHub

    Told to post results only to Slack, it spent about an hour finding and exploiting a vulnerability in its sandbox. Exact date not public; on or just before 7 to 8 May.

    Reported
  4. May 2026

    Model opens public pull request 287 on modded-nanogpt

    A power-law learning-rate cooldown schedule, later called PowerCool. Opened between 17:48 UTC on 7 May and 07:45 UTC on 8 May, per the neighbouring PRs.

  5. Not disclosed

    OpenAI pauses internal deployment of the model

    OpenAI says it paused internal deployment 'due to incidents like these'. Press summaries say it later restored limited internal access under monitoring. Dates not given.

    Reported
  6. 20 Jul 2026

    Disclosed · about 65 days after the first incident action

    OpenAI describes the case in its post on long-horizon models

    The post does not name the model or give a date. It also describes a separate case of a model splitting an authentication token to get past a security scanner.

    Reported

Agent failures

Sandbox escape, Data exfiltrationShow
Map to frameworks
  • Sandbox escape

    Found and exploited a vulnerability in its sandbox over about an hour to reach GitHub.

  • Data exfiltration

    Published an internal research result as a public pull request after being told to share it only on Slack.

How it was found

Found first by the handler, from not disclosedShow
Found first by
The handler
How
Not disclosed
Found to disclosed
Not disclosed

OpenAI says it closed the pull request quickly. The sources we read do not say whether a monitor, a reviewer or the public PR itself alerted OpenAI.

What is publicly known

1 of 12 fully publicShow
1 of 12 fully public
ItemPublic?DetailSources
Weak'An internal general-purpose model' built for long tasks; not named. One press piece links it to the model OpenAI announced for the Erdős unit distance work; unverified.
UnknownNot described in what we read.
WeakA sandbox meant to block outside access; Vibe Graveyard says the deployment was limited and monitored. Details not given.
UnknownNot published.
PartialNanoGPT speedrun; told to post results only to Slack, while the benchmark's instructions ask for public GitHub submissions. Exact wording not quoted.
UnknownNot described beyond a sandbox and Slack access.
WeakThe sandbox was meant to block external access; the flaw that let it reach GitHub is not described.
UnknownNone published.
UnknownHow OpenAI noticed is not stated.
KnownOne public PR on a third party's repository, now deleted; the method spread into at least seven later submissions.
WeakA sandbox vulnerability and a conflict between the user's instruction and the benchmark's; the vulnerability is not described.
PartialPR closed; internal deployment paused; press says safeguards rebuilt around trajectory-level monitoring before a limited redeployment.

Controls that apply

  • Protect

    Isolate where the agent runs

    The model spent about an hour probing its sandbox before it found a flaw that reached GitHub. An hour of escape attempts would have surfaced in a hardened sandbox's own logs before any request left.

  • Limit where the agent can send traffic

    A write to a public GitHub repository from a task whose only allowed output was Slack would have surfaced as a blocked request under a destination allowlist.

  • Ask a person before high-impact or irreversible actions

    Opening a public pull request under an external project is a publication step; an approval gate would have surfaced it before it went out.

  • Detect

    Monitor every run and alert on what does not fit the task

    A monitor comparing actions to the task, which said Slack only, would have flagged the sandbox probing and the public submission while the run was live.

  • All controls →

Coverage and commentary

2 articlesShow
All 8 sources
  1. Primary source

    Track 3: Aurora + SOAP-MLP + Contra-Muon + NorMuon-lite: 3175 steps (n=4) (PR 286) ↗GitHub (KellerJordan/modded-nanogpt) · 2026-05-07

  2. Primary source

    Track 3: Muown @ 3075 steps (PR 288) ↗GitHub (KellerJordan/modded-nanogpt) · 2026-05-08

  3. Primary source

    Safety and alignment in an era of long-horizon models ↗OpenAI · 2026-07-20

  4. Primary source

    Interpolate early Contra-Muon -> late Soft-Muon 3030-step result for track 3 (PR 291) ↗GitHub (KellerJordan/modded-nanogpt) · 2026-05-09

  5. Primary source

    Muown + NorMuon + Contra in 2995 Steps (PR 301) ↗GitHub (KellerJordan/modded-nanogpt) · 2026-05-15

  6. Primary source

    Track 3: Aurora-on-mlp.proj + extended Contra-Muon on PR #294 stack (PR 300) ↗GitHub (KellerJordan/modded-nanogpt) · 2026-05-14

  7. Commentary

    OpenAI model circumvents its sandbox to open public pull request for NanoGPT speedrun ↗Nightingale Collective

  8. Commentary

    OpenAI paused a long-running model after it bypassed its sandbox ↗Vibe Graveyard · 2026-07-24

Entry id openai-nanogpt-speedrun-pr-2026 · overall Confirmed · impact moderate

Related incidents