OpenAI warns ‘there may be more’ Hugging Face incidents
OpenAI and Anthropic have admitted their AI models could continue to break out of their test systems and attack third parties, as MPs grilled the world’s biggest tech companies over their ability to control their increasingly powerful systems.
Executives from the two firms told parliament on Tuesday that investigations into previous security breaches were ongoing and they were not able to rule out further discoveries of unauthorised activity.
The admissions came during a Business, Innovation, Science and Trade (BIST) Committee hearing examining whether Britain’s voluntary approach to AI safety is sufficient to protect businesses and the public.
Tom Duff Gordon, OpenAI’s vice president of EMEA policy, told MPs that the ChatGPT maker was still investigating the behaviour of its models following July’s unprecedented incident in which an AI agent escaped a controlled test and hacked the systems of Hugging Face.
“I cannot assure the committee or you, Chair, that that is it,” he said when asked whether further incidents could emerge. “There are a handful of others that I am aware of that we have notified, and there may be more to come”.
Duff Gordon said OpenAI had launched an extensive investigation into whether its models had accessed other websites without authorisation during testing.
He stressed that most of the other incidents identified involved agents retrieving information rather than breaching security systems.
The executive also apologised for delays in notifying Australian authorities after an OpenAI agent accessed government systems. “We took too long, we apologise for that”, he said, acknowledging that the company “could have and should have notified earlier”.
He also called for mandatory incident reporting rules to establish clearer requirements for when AI companies must inform authorities about unexpected behaviour.
Anthropic’s head of policy, Liam Booth-Smith, similarly acknowledged that further incidents could be uncovered.
The company published details of nine incidents affecting US organisations last week after reviewing months of testing records.
Asked whether more cases could emerge, Booth-Smith said: “I can’t give you an assurance that there aren’t”. He said Anthropic was continuing to review model behaviour, with human verification slowing the investigation.
The admissions follow mounting scrutiny of AI developers after a series of incidents in which increasingly autonomous systems accessed external websites or behaved in ways their creators had not intended.
OpenAI admits safety setback
OpenAI also confirmed that it had withheld the release of its latest GPT model after internal testing revealed a deterioration in its ability to follow intended safety constraints.
“We actually saw a regression in the alignment of that model”, Duff Gordon said. “Because of that, we decided not to deploy it”.
The executive said the model had performed worse than its predecessor in following instructions in line with human values.
Elsewhere, Google Deepmind’s Owen Larter told MPs he was unaware of further incidents involving its models, while Meta’s Rob Sherman said the company was not aware of cases similar to those already disclosed.
The hearing followed warnings from former AI researchers earlier on Tuesday that existing safety evaluations could fail to detect dangerous behaviour.
Steven Adler, a former OpenAI safety researcher, warned that advanced models could recognise when they were being tested and conceal unwanted behaviour.
“The system could be playing possum until a moment when they put it out into the world, and then it takes a turn for the worse”, he said.
Daniel Kokotajlo, another former OpenAI researcher, argued that companies should remain liable when their systems act beyond their control.
“You definitely shouldn’t let the companies get away with evading responsibility for what their AIs do just because their AIs went rogue”, he said.
The evidence comes a day after AI minister Kanishka Narayan warned that a major breach of Britain’s cyber defences was likely to be inevitable.
Britain’s AI Security Institute is already investigating the Hugging Face incident, while ministers face growing pressure to introduce mandatory testing and reporting requirements for advanced AI systems.