OpenAI said its AI agents leaked 53 images belonging to ChatGPT users, marking the latest in a growing list of incidents involving Artificial Intelligence acting beyond their intended limits.
Since then, more than 15 incidents involving OpenAI agents have been reported by the company or outside researchers, Reuters reported. OpenAI has also reported several previously unknown cases and introduced a new system for reporting AI safety incidents. The company also said that many incidents reported by outside researchers overlap with cases already under its investigation.
The incidents have included spam-like activity on websites and attempts to exploit software weaknesses.
The company said it would be transparent “even when significance is uncertain. OpenAI said it is focusing on the most serious cases as its investigation continues.
This comes about two months after OpenAI disclosed that its AI agents accessed the systems of AI repository Hugging Face during an internal test. Because they were looking for reliable sources of publicly available information, openAI said its research models accessed some government, university and public-agency websites.
It sent questions to an outside chatbot before the company stopped the training run about 2.5 hours later. OpenAI said its monitoring system detected the activity within 15 minutes, and a human reviewed the alert three minutes later. The company said it found no evidence of unauthorised access, hacked accounts or security breaches, Reuters reported.
Because the company uses some anonymised ChatGPT user data to train its models, openAI’s AI agents could access the images. This came after the viral 1980s AI photo trend on Instagram where users used ChatGPT to transform their pictures into retro photographs resembling vintage family albums, classic studio sessions, or old Bollywood stills The company did not immediately reveal whether the leaked images were AI-generated or showed real people. It also did not disclose when exactly the images were originally posted. Enterprise data is not used for training, while consumer users can choose not to allow their data to be used for training. OpenAI had found about two dozen cases of agents behaving unexpectedly by mid-September. The number has continued to increase as the company reviews its internal records. It also paused training and testing of its most capable models when they use external tools. Billionaire Elon Musk also supported Amodei, writing, “Dario is right.
Most of the images have been removed as the company says it is working with hosting providers to take down the remaining ones, according to a Reuters report. The AI giant said it removed details such as names, contact information and metadata from data before using it for training. People familiar with the company’s practices told Reuters that some personal information may still remain. OpenAI said the investigation could take months. Earlier, OpenAI said an AI agent being trained in a supposedly internet-free environment found a way to access the public internet and contact another chatbot. The agent used a gap in the DNS filtering of its training system, according to OpenAI. The company says it has since added two separate security measures that would have blocked the access. The agent sent at least 20 questions to the outside chatbot, including “What is the capital of France”, Bloomberg reported. OpenAI said it would not restart training of that particular model. AI research group Transluce said agents believed to be linked to OpenAI tried but failed to access a US Department of Education website dealing with civil rights. OpenAI CEO Sam Altman backed the call and said the industry needed to “pace the frontier”. Meta CEO Mark Zuckerberg, though, said that AI companies had their own reasons and ability to put safety measures in place, while Google DeepMind’s Demis Hassabis said Amodei’s proposal pointed in the right direction but that its details needed to be worked out.
Earlier this month, Anthropic CEO Dario Amodei called for the industry to slow the pace of AI development, saying rapid progress could make the technology harder to understand and control. On Friday, OpenAI said its AI models accessed information from the websites of the US Securities and Exchange Commission and the US Census Bureau during research and training.
