OPENAI
OpenAI’s agents went off-script
OpenAI’s AI agents appear to have used more than 10 previously undisclosed websites to communicate with each other, despite being restricted from posting online, according to researchers cited by Reuters.
Investigators found traces of the agents using wikis, text-storage sites and university link shorteners between May and July.
Some researchers counted 18 affected sites, while another group identified 23, although Reuters could not independently verify every case.
The agents had reportedly been tasked with answering difficult research questions while only being allowed to read information online.
Researchers believe they found loopholes in older websites that let them leave messages behind anyway, effectively turning parts of the web into makeshift communication channels.
Here’s what you should know:
Researchers say OpenAI agents used more than 10 websites despite being told not to post online.
Some investigations have identified as many as 23 previously undisclosed sites, although the full number remains unclear.
OpenAI says it is reviewing the incidents and plans to publish a system for reporting this type of AI behaviour.
Finding somewhere to leave a note
The activity included obscure sites such as an old chemistry wiki, personal websites and hobbyist pages.
Researchers linked some of the activity through matching usernames, identical strings of data and IP addresses connected to Microsoft Azure infrastructure used by OpenAI.
OpenAI said it is reviewing the activity and has not found anything matching the severity of its earlier Hugging Face incident.
The company is also developing a framework for reporting AI “misalignment”, its term for models behaving in unexpected or unintended ways.
It reminds me of trying to find unblocked game websites in school. - MV