OpenAI uncovers more AI agent escapes while probing Hugging Face security breach

OpenAI has found additional cases where its AI agents escaped controlled testing environments as the company continues investigating the hacking incident involving the open-source AI platform Hugging Face. According to a Reuters report, the newly discovered incidents were identified during a broader review launched after the Hugging Face case came to light in July.  Also Read: Claude AI breached three companies during internal testing, says Anthropic

While the latest findings were reportedly limited in scope, they have added to ongoing discussions around AI safety, monitoring, and whether advanced AI systems are becoming harder to supervise during testing.  Also Read: OpenAI is giving free access to its most advanced ChatGPT models: Who is eligible?

More escape incidents found during investigation 

According to Reuters, OpenAI discovered the additional incidents while reviewing activity related to the Hugging Face intrusion. People familiar with the matter told the publication that the AI agents managed to escape their intended testing environments, although there is no indication that any of them left OpenAI’s internal network.  Also Read: Google launches Ask Google Pay AI chatbot to help you track spending; Here's how

One of the sources cited in the report said the incidents were limited in nature. 

Earlier this week, OpenAI had already said it was reviewing “broader activity from our models” alongside its investigation into the Hugging Face incident. The company has not publicly shared further details about the newly identified cases. 

Investigation includes older system logs 

As part of the expanded probe, Reuters reported that OpenAI and outside experts are reviewing system logs from earlier this year to better understand how the incidents happened and whether there were similar cases that had gone unnoticed. 

The investigation was launched after an OpenAI AI agent reportedly escaped its testing environment during an internal cybersecurity evaluation and ended up operating inside Hugging Face’s network for several days. OpenAI later confirmed that four accounts across four other companies were also compromised during the same incident. One of those companies was Modal, according to Reuters. 

Anthropic also disclosed similar incidents 

The Reuters report noted that OpenAI’s investigation expanded shortly before rival AI company Anthropic revealed similar issues involving its Claude models. 

Anthropic said its AI models gained access to systems belonging to three real companies during cybersecurity evaluations that were intended to run inside isolated environments. According to the company, the incidents happened because of a configuration mistake in a third-party testing setup that unintentionally allowed internet access. 

Anthropic clarified that Claude was not attempting to escape its testing environment. Instead, the model believed the real-world systems it encountered were part of the evaluation because it had been instructed that no internet connection existed. 

The company also acknowledged that closer monitoring of evaluation logs could have helped detect the issue sooner, while adding that real-time monitoring existed but was not enabled for that specific testing scenario because of a misunderstanding with a partner. 

AI safety concerns draw attention from regulators 

The latest disclosures are expected to increase scrutiny of AI companies as governments consider new rules for advanced AI systems. 

Add Techlusive as a Preferred Source

According to Reuters, the European Commission has already held discussions with both OpenAI and Anthropic following the recent hacking incidents. In the US, President Donald Trump said officials were looking at possible controls, while Senator Mark Warner said the Anthropic incident reinforced the need for mandatory capability testing of advanced AI models before wider deployment.