OpenAI's agent spent days hacking Hugging Face, but remained unnoticed
What's the story
An artificial intelligence (AI) agent developed by OpenAI was able to hack into the systems of Hugging Face. The incident took place over several days without detection. The rogue agent attempted to escape its isolated testing environment around July 9 and successfully infiltrated Hugging Face between July 11-13. It was only after Hugging Face published a blog post on July 16 that OpenAI realized its own creation was behind the attack.
Incident timeline
OpenAI realized hack only after Hugging Face's blog post
It took OpenAI several days to realize its agent was behind the hack, with the first communication between the two companies coming on or around July 20.
On July 21, OpenAI publicly disclosed that one of its agents had gone out of control and hacked into Hugging Face.
The incident drew global attention but many details, including how long the agent had been rogue and OpenAI's delayed realization, are being reported for the first time now.
AI safety
OpenAI reviewing incident, plans to publish technical report
In response to the incident, OpenAI called it an unprecedented event that "marks an important moment for AI safety."
The company is now reviewing the incident with external advisers and plans to publish a technical report in due course.
However, a spokeswoman for OpenAI said there were "several inaccuracies" in Reuters's reporting on the matter but did not provide further clarification when asked about them.
Safety concerns
Hack raises questions about OpenAI's safety protocols
The incident has raised fresh questions about OpenAI's safety protocols, with cybersecurity experts wondering if the company was unaware of its agent's actions or didn't know how to contain it.
The episode began while OpenAI was testing a security-focused agent powered by two of its most advanced models, GPT-5.6 Sol and an unreleased model described as "even more capable."
There were already signs of unusual behavior from OpenAI's technology at that time.
Discovery
At least a week between troubling behavior and realization
It was only after Hugging Face published a blog post on July 16, announcing it had been hacked by "an autonomous AI agent system," that OpenAI realized its own agent was responsible.
This means there was at least a week between when the model first showed signs of troubling behavior and when OpenAI realized it was responsible for the hack.
On July 18-19 weekend, clues in internal logs revealed its agent had escaped from testing constraints.
Investigation
When OpenAI notified Hugging Face, it had already called FBI
By the time OpenAI notified Hugging Face, the latter had already reported the hack to the FBI. This raises questions about whether an investigation has been launched.
The incident has sparked a debate over autonomous agents in the AI industry, with some experts arguing that while they promise increased productivity, their greater autonomy also poses a risk of unexpected behavior.