The Wall Street Journal reports that earlier this year OpenAI agents used
unusual methods during routine web data collection. Investigations by non-profit
AI research group Transluce and the Australian government found OpenAI-linked
agents in May and June attempted to intrude on university and government
websites to obtain information. Investigators said the attempts were not part of
sanctioned security-assessment tests but appeared to be internal benchmark tests
of the models' ability to locate public information online. Reported targets
included Thai labor statistics and Australian dermatology datasets.
