← AI Safety

Spaces

As of September 2026

A "Space" is Hugging Face's name for a small hosted web app — usually a demo that lets you try a model in a browser without installing anything. Spaces are also where an earlier, unrelated Hugging Face security incident happened, back in 2024, which is worth knowing about separately from the July 2026 attack this site is mainly about.

What a Space is made of A vertical diagram. Two boxes side by side, a Space repository containing app code, and hardware such as a CPU or GPU tier, combine with a plus sign between them. An arrow leads down to a single box representing the running app, publicly reachable at its own web address, which holds any secrets such as API keys as hidden environment variables. Space repo app code (e.g. app.py) + Hardware CPU or GPU tier you choose Running app live at its own web address anyone can open and use Spaces secrets API keys held as hidden environment variables Visitors see the app; they don't see the code, the secrets, or the underlying hardware.
A Space is a repo, plus hosting hardware, running continuously as a web app — closer to a small website than a downloadable file.

What a Space is

Where a model repo or a dataset repo is mainly something you download, a Space is something you run — or rather, something Hugging Face runs for you. A Space is a repository of app code (often a short Python script using a framework like Gradio or Streamlit) paired with a slice of Hugging Face's own hardware, kept running continuously so anyone with the link can open it in a browser and interact with it live. It's the closest thing on the Hub to a website: researchers use Spaces to let people try a new model without installing it; companies use them for public product demos; hobbyists use free-tier Spaces for personal projects.

Free and paid hardware

A Space runs on hardware Hugging Face provides, not the visitor's own computer, and that hardware is billed to the Space's owner rather than metered per visit. There is a free tier suitable for small demos, and paid tiers with more powerful hardware for anything larger. The owner picks that tier explicitly — it isn't automatic — which is why "hardware you choose" appears as its own piece in the diagram above, separate from the app code itself.

Public and private Spaces

Like model and dataset repos, a Space can be public, with its code and its running app visible to anyone, or private, restricted to its owner or an organization's members. A private Space is the usual choice for an internal tool or a demo that isn't ready to share. "Private" here controls who can view the code and use the running app; it's a separate setting from Spaces secrets, which stay hidden from viewers of a Space regardless of whether the Space itself is public or private.

Spaces secrets

Because a Space is running code continuously, it often needs its own credentials — an API key for a paid service it calls, a token that lets it read a private model, and so on. Hugging Face lets a Space's owner store those as "Spaces secrets": values set once, kept hidden from anyone viewing the Space's code or its visible output, and injected into the running app as environment variables. Visitors to the Space never see them; the app has to keep them out of anything it prints or displays.

May 2024: unauthorized access to Spaces secrets

In May 2024, Hugging Face detected unauthorized access to its Spaces platform and warned that secrets stored by Spaces owners may have been improperly accessed. The company revoked and rotated keys on its side, and advised anyone running a Space to refresh their Hugging Face tokens and rotate any secrets stored in that Space, on the reasoning that a secret which may have been exposed should be treated as compromised rather than assumed safe. It's a useful habit independent of any one incident: the same "rotate rather than hope" advice applies to the July 2026 attack too, covered on how to protect yourself.

Did the July 2026 attack touch Spaces?

Based on Hugging Face's own account of the July 2026 intrusion, no. Hugging Face's technical post-mortem states plainly that no other customer-facing models, datasets, Spaces, or packages were affected, and that its software supply chain was checked and found clean. The 2026 attackers reached the Hub through its dataset-processing pipeline, not through Spaces — the mechanics of that entry point are covered on the Hub page and how the attack worked. That doesn't make Spaces secrets a non-issue going forward — the May 2024 incident above shows they've been a target before — it just means this specific 2026 attack, as Hugging Face has described it, is a separate story from Spaces.

Sources

  1. Hugging Face docs, "Hub" — https://huggingface.co/docs/hub/en/index
  2. TechCrunch, "Hugging Face says it detected unauthorized access to its AI model hosting platform," 31 May 2024 — https://techcrunch.com/2024/05/31/hugging-face-says-it-detected-unauthorized-access-to-its-ai-model-hosting-platform
  3. Hugging Face Security Team, "Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident," 27 Jul 2026 — https://huggingface.co/blog/agent-intrusion-technical-timeline
  4. Hugging Face Security Team, "Security incident disclosure — July 2026," 16 Jul 2026 — https://huggingface.co/blog/security-incident-july-2026