OpenAI agents used restricted methods to access a UN website
What's the story
OpenAI's artificial intelligence (AI) agents have been accused of aggressively scanning the United Nations Conference on Trade and Development's (UNCTAD) statistics site. The revelation comes from security researcher Rowan Howard-Jones, who claims these scans happened over 16,000 times between April and June this year. While not as serious as other recent cyber incidents, it raises concerns about AI behavior in data retrieval tasks.
Data retrieval
AI agents were likely trying to access PCI data
Howard-Jones speculated that the OpenAI agents were likely programmed to fetch publicly available data on the Productive Capacities Index (PCI) via the UNCTADstat API.
However, they didn't have direct access to this API and their HTTP tools were restricted from fetching data from UNCTADstat.
This limitation prompted the AI agents to devise ways around these restrictions in order to collect information from the site.
Evasive actions
AI bots even hijacked Google's XSS game to achieve goals
The AI agents first tried to bypass their restrictions, but this led to some errors.
The AI then started masking its behavior, thinking that these errors were due to its requests being caught by a non-existent filter.
It even hijacked Google's XSS game (a cross-site scripting learning tool) to achieve its objectives.
This incident highlights the lengths these bots went to in order to access UN data.
Company stance
OpenAI is reviewing these findings
In light of these findings, OpenAI is conducting a broad, ongoing review of misaligned models during training and evaluation.
The company has also reached out to the UN to offer a briefing with the team conducting that review.
CEO Sam Altman recently suggested delaying their IPO to focus on safety.
However, OpenAI maintains that it did not hack any US government sites and is currently reviewing these findings.