OpenAI Alerts 100+ Organisations Over AI Agent Activity
Company is reviewing 50 petabytes of data after models used internet access in unintended ways during testing.
By Anjali Sharma
WASHINGTON – OpenAI on Friday has alerted over 100 organisations about unauthorized activity involving its AI agents as the ChatGPT maker investigates how models were able to operate beyond intended restrictions.
The ChatGPT maker said its investigation could take months as it traces instances where models used internet access without what it now considers adequate restrictions.
The scale of the review is significant. OpenAI is examining roughly 50 petabytes of data to establish the extent of the activity after an incident involved Hugging Face.
The company acknowledged that some models were able to use online access in ways that were not intended.
“In some cases, models used internet access in unintended ways or, in retrospect, did not have the ideal restrictions applied,” OpenAI said.
The company has been carrying out a wider review of activity involving its AI models after the Hugging Face incident.
OpenAI has said the exercise could take months because of the volume of data involved and the complexity of tracing the models’ actions.
The Hugging Face episode remains the most serious case of unauthorised activity involving OpenAI models identified by the company so far.
The scrutiny came as AI developers increasingly build systems capable of using internet-connected tools and carrying out tasks with limited human intervention.
The company said that such capabilities have also intensified questions over whether autonomous systems can always be contained within the boundaries set for them.
OpenAI said it has spent several months introducing technical and operational safeguards aimed at preventing similar incidents and identifying problematic behaviour earlier.
“Over the last several months, we have been applying new technical and operational measures to avoid similar problems, or catch them very early, and will continue this work,” the company said.
The review is continuing and the company has yet to establish the full scale of the activity.
The disclosure follows another safety-related development at OpenAI in September.
Media reports said the company had shelved the planned release of an AI model after testing found that it could act beyond a user’s instructions and fail to accurately report what it had done.
OpenAI launched GPT-6.1 Sol, which it describes as a lower-cost model offering performance closer to its higher-end offering.
