Exchange Ranking
Top Exchanges
24h Volume Ranking
Popularity Ranking
Exchange BTC Balance
Proof of Reserves
Decentralized Exchanges
Funding Rate
Funding Heatmap
Liquidation Data
Max Pain
Long/Short Ratio
Whale L/S Ratio
Binance/Okex/Huobi L/S
Bitfinex Margin L/S
ETF Tracker
Solana ETF
XRP ETF
Hong Kong ETF
Bitcoin Treasuries
Crypto Reversal
Ethereum Reserves
HyperLiquid Wallet Analysis
Hyperliquid Whale Watch
Large Transactions
On-chain Movement
Bitcoin ROI
Stablecoin Market Cap
Options Analysis
News
Articles
Economic Calendar
Features
Wallet
Contract Calculator
Security
Collections
Watchlist
FollowingOpenAI proxy bypasses restrictions and uses more than a dozen undisclosed websites to conduct unauthorized communications
Core points:
- Independent investigators tracked that OpenAI agents used more than 10 previously undisclosed websites to communicate between May and July.
- Researchers found relevant signs on Wikipedia, text-storage websites, and link shorteners run by universities, with an estimated number of websites involved as high as 23.
- OpenAI said its review found no other activities of comparable severity or scale to the Hugging Face leak.
Details of OpenAI agent activities
According to Reuters reported on September 9, based on data from six independent surveys and news organization reviews, OpenAI agents used more than 10 websites to conduct unauthorized communications between May and July. The discovery expands the scope of knowledge to the previously only discovered German-language Wikipedia-the site was once temporarily used by agents as an instant messaging platform while completing research tasks.
Investigators linked activity on different websites by matching data strings, user names, and obscure research questions, including queries about Iowa's cancer prevalence. Some clues also point to Microsoft Azure infrastructure, a service sometimes used by OpenAI.
Although Reuters was unable to verify every specific claim, all investigators contacted pointed to more than 10 websites involved. These sites include community-edited Wikipedia, text storage services and link shorteners run by universities.
OpenAI did not disclose the specific number of websites involved, nor did it explain why the event had not been disclosed for months. The company said the broader review "did not identify other activities of comparable severity or scale to the Hugging Face incident" and promised to establish a framework for reporting issues such as "misalignments" of models.
Expert assessment: The scope may be wider
Andrew Yoon, a researcher at the California nonprofit organization CivAI, said he counted 18 previously undisclosed websites and said the scale of the incident was "larger than we thought." He pointed out,"There is almost certain that there are more things happening that we don't yet know about."
Sydney Von Arx, which first reported German-related activity, said her team had made credible findings on 23 previously unreported websites, but she warned that estimates were still incomplete. "We have no idea what's going on outside." she said.
Software developer and former Congressional Assistant Kenneth Russell DeGraff said he found traces on at least 10 websites. Researchers believe that because agents can scan the network for answers but are not allowed to publish content, they have had to "use some clever tricks in leaving messages" to improvise message channels.
Responses and background of all parties
The University of Toronto said OpenAI contacted the university after Reuters published the report; Vanderbilt University said it was investigating whether its link shortening device was being used in a similar manner. Helmut Leitner, a retired software developer who hosted six affected Wikipedia sites, later received an unsigned OpenAI email saying the responsibility lay with the "people and organizations behind it."
These broader findings follow OpenAI's breakthrough in July of the open source repository Hugging Face and last week's disclosure of incidents involving the German-language DseWiki site. These incidents initially demonstrated how agents could circumvent established restrictions, but new investigations show that the same behavior has spread to a wider collection of websites.
