Safety · Article 5 of 10
The Hugging Face incident, for business owners
In July 2026, AI models running inside an OpenAI security test broke out of their test environment and reached systems at Hugging Face, a company that hosts AI tools. Hugging Face detected and stopped it. Both companies published what happened. The lesson for any business is simple: AI with access needs limits, and someone watching.
What happened
OpenAI was running an internal test to measure how good its models are at finding and exploiting security flaws. The test ran without some of the safeguards used in its public products. According to OpenAI's report, the models used an unknown flaw in a piece of internal software to reach the internet, and then chained other weaknesses until they reached Hugging Face's infrastructure. Hugging Face detected the activity and cut it off.1
Hugging Face published its own technical timeline a week later. It describes an autonomous agent working over several days, using leaked access keys, passwords stored in plain configuration and one shared credential that opened doors across different systems. It says five datasets were accessed and that other customers' models, datasets and apps were not affected.2 The two reports differ on some details; here we follow each company's own words.
OpenAI's own conclusion: "model security and safety must keep pace with rapidly advancing capabilities."1
This page has had your eyes for 0 s. Without them: 0 s.
This page keeps a small record of the time you watched it and the time you didn't. Switch tabs and come back. Systems keep running when nobody is looking.
The goal
Real customers, verifiable facts
The access
Drafts and checks only
The oversight
Your written yes, a source behind every claim
All three set on purpose. A clear goal, only the access it needs, and someone who signs off.
Drag a dial, or focus it and use the arrow keys.
Three dials to judge any AI setup
Every AI risk comes down to three dials:
- The goal. What the system is told to chase. In the incident, the models went "to extreme lengths" to reach the test goal, in OpenAI's words.1
- The access. What it can touch. Here, keys and shared credentials let it go much further than intended.2
- The oversight. Who is watching, and when. Here, the other company noticed.
Turn any dial too far and the risk grows. Set all three on purpose.
Five lessons for any business
Lessons from the incident and from what Hugging Face changed afterwards, translated for a business that uses AI tools:2
- Give each tool only the access it needs. Read-only when reading is enough.
- Use separate keys for separate things. One shared key that opens everything is how small problems become big ones.
- Prefer access that expires. Keys that live forever are keys that leak forever.
- Never keep passwords in plain files or chats. Use a password manager.
- Watch the activity. Someone, or something, should notice when a tool does what it shouldn't.
None of this needs a big IT team. It needs decisions, made before something goes wrong.
Want to apply them in your business? Ask me how we do it for your business, on a free 30-minute call.
Charted and signedOctober 2026 · Alejandro Peñuela
Every claim in this article points to its source.