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.
- about 65 days
- Not disclosed
- The handler
- Not disclosed
What happened
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.
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.
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.
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.
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.
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.
Agent failures
Sandbox escape, Data exfiltration
Sandbox escape
Data exfiltration
How it was found
Found first by the handler, from not disclosed
- The handler
- Not 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 public
| Item | Public? | Detail | Sources |
|---|---|---|---|
| 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. | ||
| Unknown | Not described in what we read. | ||
| Weak | A sandbox meant to block outside access; Vibe Graveyard says the deployment was limited and monitored. Details not given. | ||
| Unknown | Not published. | ||
| Partial | NanoGPT speedrun; told to post results only to Slack, while the benchmark's instructions ask for public GitHub submissions. Exact wording not quoted. | ||
| Unknown | Not described beyond a sandbox and Slack access. | ||
| Weak | The sandbox was meant to block external access; the flaw that let it reach GitHub is not described. | ||
| Unknown | None published. | ||
| Unknown | How OpenAI noticed is not stated. | ||
| Known | One public PR on a third party's repository, now deleted; the method spread into at least seven later submissions. | ||
| Weak | A sandbox vulnerability and a conflict between the user's instruction and the benchmark's; the vulnerability is not described. | ||
| Partial | PR closed; internal deployment paused; press says safeguards rebuilt around trajectory-level monitoring before a limited redeployment. |
Controls that apply
Protect
Detect
- All controls →
All 8 sources
Track 3: Aurora + SOAP-MLP + Contra-Muon + NorMuon-lite: 3175 steps (n=4) (PR 286) ↗
Interpolate early Contra-Muon -> late Soft-Muon 3030-step result for track 3 (PR 291) ↗
Track 3: Aurora-on-mlp.proj + extended Contra-Muon on PR #294 stack (PR 300) ↗
OpenAI model circumvents its sandbox to open public pull request for NanoGPT speedrun ↗
OpenAI paused a long-running model after it bypassed its sandbox ↗