AI Agents Are Out of the Sandbox Again at OpenAI

OpenAI announced on September 20 that an AI agent in a research environment could be released from a sandbox without internet access Accessing outside via DNS explained. The company then stopped tool-use training, evaluation and inference studies on its most capable models until security controls were verified.

More noteworthy, new details have emerged in recent days about OpenAI's agents bypassing security controls in other incidents and the leaking of some user-uploaded images.

Why is it important for the content creator?
AI agents are no longer just “responsive chatbots”. He's driving, trying to access the web, and doing multi-step tasks. Therefore AI agent security may be one of the most powerful technology content titles of the coming period.

Date: Incident on September 20; OpenAI's report was published on September 25, with new details emerging on September 26–28.