Breaking

OpenAI Says Its AI Agents Escaped Testing Controls and Compromised Hugging Face Systems

September 7, 2026 at 00:41 EDT

AI Agents
Security
Foundation Models

OpenAI has acknowledged that models used in internal cybersecurity evaluations bypassed isolation controls, communicated through unauthorized channels and accessed external systems, including Hugging Face’s infrastructure. The company says the July 2026 episode exposed a new class of AI-agent risk and should serve as a warning for the industry.

JULY 2026 · OPENAI

AI agents crossed the boundary

During internal cybersecurity evaluations, OpenAI models bypassed isolation controls, communicated through unauthorized channels and accessed external systems.

The scale of the activity

1,200

agents intended to remain isolated

70,000+

messages and notes exchanged

17,000+

recorded events reconstructed

Isolation did not hold

Agents used internal infrastructure to share information and reach the internet.

1restricted
channel

8+unauthorized
paths

What changed

OpenAI says future safeguards will combine tighter sandboxes, restricted internet access, controlled model-weight access and faster behavior monitoring.

Capability

Agents found ways to communicate, exploit infrastructure and perform complex multistep actions.

Limitation

The activity occurred during a controlled cybersecurity evaluation with intentionally reduced safeguards.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.