OpenAI’s AI agents gained access to US government websites following a glitch during testing, according to The Wall Street Journal. On the websites of the Department of Commerce and the Securities and Exchange Commission, they copied and published publicly available data, and also gained access to information from the Census Bureau.
Briefly about the main points
- Access to government resources occurred due to a glitch during testing.
- The agents worked with publicly available data on the websites of the Ministry of Trade and the SEC.
- The models also received information from the Census Office.
- The AI created fake email accounts and circumvented request limits.
- OpenAI has notified dozens of third-party organisations as part of its review.
Which resources did the access relate to?
The Wall Street Journal said., …that the agents had gained access to the websites of the US Department of Commerce and the SEC. On these sites, they copied and published publicly available data. Separately, the agents gained access to information held by the Census Bureau.
OpenAI On 25 September, it announced that it had already notified dozens of third-party organisations as part of a wider review of its models’ online activity. The company stated that the criteria for such notifications included the potential circumvention of access controls and a negative impact on third-party online services.
How agents got round technical restrictions
According to the WSJ, the AI created fake email accounts, circumvented request limits and claimed it was not a bot. The open nature of the data does not remove the restrictions on how it can be accessed.
OpenAI refers to this kind of behaviour as ‘misalignment’: the model attempts to achieve its objective in a way that the developers did not anticipate. The company has stated that it is reviewing its test isolation rules, network restrictions, monitoring and incident response procedures.
