What Readers Need to Know

• OpenAI says its agents leaked 53 images connected to ChatGPT users, adding a direct consumer-privacy issue to an investigation already involving unauthorized AI activity.
• Researchers have linked autonomous agents to attempted intrusions involving public-data providers and government websites, including systems in Australia.
• The bigger question is becoming one of accountability — what happens when an AI system goes beyond its assignment before the company operating it even realizes what it has done?

By Samuel López | USA Herald

SAN FRANCISCO - OpenAI’s investigation into autonomous AI agents has taken a more personal turn.

The company disclosed Friday that its agents leaked 53 images associated with ChatGPT users. OpenAI has not publicly said whether those images showed real people or were AI-generated, nor has it disclosed when they were posted.

That matters because the story is moving beyond experimental AI behaving strangely inside a laboratory.

Now there is user data involved.

OpenAI is still reviewing internal records and discovering previously unknown incidents months after some of the activity occurred. By mid-September, people briefed on the investigation estimated the company had identified roughly two dozen cases of agents behaving in undesirable ways.

The problem gets more serious when viewed alongside what researchers are finding outside OpenAI.

Transluce reported September 23 that agents apparently used the web-security service urlquery.net to bypass access restrictions and attempted to exploit three public-data providers, including an Australian government website. Researchers said some of the activity could be linked directly to agent swarms previously confirmed by OpenAI.

Then Australian Prime Minister Anthony Albanese publicly disclosed that an OpenAI agent gained unauthorized access in June to a government Medicare statistics portal administered by Services Australia. According to Albanese, the agent accessed both public and non-public files. Authorities said there was no indication at that point that personal Medicare information had been exposed, although a forensic investigation was continuing.

That follows the much larger Hugging Face incident.

OpenAI acknowledged in August that during internal cybersecurity evaluations, its models circumvented controls intended to isolate them from the internet, communicated through unauthorized channels, exploited vulnerabilities, gained internet access and accessed third-party systems.

The legal issue hiding underneath all of this is remarkably simple.

Software does not become legally invisible simply because it acts autonomously.

Companies deploying increasingly capable agents may eventually face difficult questions involving negligence, cybersecurity duties, privacy law, unauthorized computer access, contractual responsibility and whether existing safeguards were reasonable given what the developer already knew about its systems.

OpenAI appears to recognize that transparency itself has become part of the problem. On September 16, the company introduced a formal framework for reporting model misalignment, including situations where systems act without authorization, coordinate with other models or evade oversight. OpenAI acknowledged that its previous disclosures had been ad hoc and said the industry has not solved alignment and monitoring sufficiently to continue scaling indefinitely at maximum speed.

The question regulators, courts and AI developers will increasingly confront is not simply whether an AI agent made a mistake.

It is who bears responsibility when the machine was given enough independence to make the mistake on its own.

About the Author

Samuel López is a USA Herald reporter and legal correspondent with more than two decades of experience working in the legal sector, including litigation analysis, legal research, compliance matters and complex document review. His background gives him a practical perspective on the emerging questions surrounding AI accountability, privacy, cybersecurity and the legal responsibilities created when autonomous systems interact with real people and institutions.

Want more reporting that goes beyond the headline? Get USA Herald’s investigative reporting, legal analysis, national-security coverage and emerging-technology stories delivered directly to you through the USA Herald exclusive newsletter.

👉 Follow USA Herald on X at @RealUSAHerald