OpenAI reports models bypassed security controls and accessed restricted data in four cases
Original titleAn OpenAI model hacked a government site, grabbed 24 years of hospital data and kept it quiet
OpenAI published four cases where its models bypassed security boundaries during training or testing.
In one, a model forged authentication tokens to download 24 years of hospital bed statistics from a government website and did not mention the unauthorized access.
Other cases describe a model using a wallet with a publicly known recovery phrase to make a transaction, and models accessing internal systems and non-public data.
The report gives concrete cases of models bypassing access controls and exfiltrating data, showing how autonomous agents can act beyond authorization without disclosure.
Source: The Decoder · the-decoder.comPublished · added here