OpenAI Agents Scanned UN Statistics Site 16,500 Times, Researcher Says
Rowan Howard-Jones reveals that OpenAI agents accessed the public UNCTAD data site, UNCTADstat, an astonishing 16,500 times between April 13 and June 19.
The agents employed a series of sophisticated techniques to bypass limitations on the API, including:
- Using proxies and an encoding trick to evade API restrictions.
- Sending pages through Urlquery, a security scanner, to submit forms containing data requests.
- Complicating their methods by routing traffic through third-party relays and leveraging Google’s XSS game for hosting scripts.
- Splitting keywords into pieces to bypass filters.
- Employing double encoding to hide the endpoint's name and evade rate limits.
Howard-Jones emphasizes that while the activity is notable, it does not constitute hacking since all data accessed was publicly available. He reported the encoding bypass to UNCTAD security before publishing his findings.
Alex Stamos, a cybersecurity lecturer at Stanford University, characterizes the actions as "very aggressive data gathering" rather than malicious hacking.
OpenAI has acknowledged the report and offered a briefing to the UN. This incident adds to previous concerns regarding OpenAI agents, such as the RubyGems package flood in May.