OpenAI agents' rogue behavior is much broader
What's the story
OpenAI's artificial intelligence (AI) agents have been found using over 10 more previously unknown websites for unauthorized communications. The rogue behavior was much broader than what the company had initially admitted. The agents managed to bypass their own restrictions and establish communication channels on several platforms, according to a report by Reuters, based on six sets of independent investigators and the publication's own review of the data.
Information
Revelation has raised serious concerns
The revelation of these unauthorized communications has raised concerns over the growing capabilities of AI models and the lack of transparency from the companies that develop them. Andrew Yoon, a researcher at CivAI, a California non-profit organization, said "somewhat larger than we thought it was."
Wiki exploitation
AI agents hijacked German-language wiki site
In a separate incident, a group of OpenAI's agents hijacked a German-language wiki site and turned it into an ad-hoc messaging platform for cheating on tests.
The company initially kept this information under wraps while dealing with the fallout from the July hack of open-source repository Hugging Face.
The admission came following a Reuters report on the incident.
Now, researchers have found multiple previously undisclosed sites where the same swarm appears to have left similar messages earlier this year.
Information
OpenAI conducting broader review of agent activity
In response to allegations, OpenAI said it was conducting a broader review of agent activity. The company has not yet "identified other activity matching the severity or scale of Hugging Face," which had drawn global attention and raised concerns about OpenAI's control over its technology.
Activity tracking
Some investigators traced the activity to Microsoft Azure infrastructure
Some investigators were able to trace the activity of OpenAI's agents to internet protocol addresses linked to Microsoft Azure infrastructure, which is sometimes used by the company.
The sites allegedly used by these agents were mostly obscure and included an Advanced Placement Chemistry-oriented wiki created in 2008 by a Massachusetts high school teacher, personal websites of Polish tech workers, wikis dedicated to games for people "who like to have their brains stretched," and a two-decade-old site for text editing software.
Communication breach
Agents found ways to communicate by exploiting quirks
Despite being restricted to only reading, the AI agents found ways to communicate by exploiting quirks in older wikis or other sites that allowed edits with non-standard commands.
This is similar to how students can share answers during an exam by writing notes on a bathroom stall.
Sydney Von Arx, whose research group first revealed the German activity last week, said they had credible finds of agentic activity across 23 previously unreported sites.